We publish the best ideas in AI and education, from the people doing the work.

If you know him, you are probably wondering when he even has the time.

spent the last year helping lead some of ASU’s AI efforts and shared 1,001 resources for faculty, for governance, for teaching and learning, and for personal use in his daily LinkedIn posts.

Still, I am completely unconcerned that he is busy, so I asked him if he would write the foreword to Transparency & Reflection: Navigating student + AI work in higher education

with chapters from , , , and

Despite his competing commitments, Adam agreed.

He did our first early read of the manuscript and offered to set the stage for us.

Here’s how he kicks off the book for us:

Introducing this volume is a challenge.

It is a challenge because the book is deeply grounded in dialogue. That dialogue is mapped throughout in the authors’ direct responses, challenges, and critiques of one another’s thinking, and it spans months—years, really, I think—of distributed conversation on LinkedIn (yes, LinkedIn), the space where I first met and engaged with these authors and their work. Trying to capture that dialogue made me feel like one of Borges’ cartographers in “On Exactitude in Science,” expanding my map further and further until it risked becoming a point-for-point replica as large as the terrain itself.

Introducing this volume is also a challenge because of the moment that drove the book’s genesis. Many educators are reaching for assessment lifelines now that AI has made the production of artifacts so frictionless, and, in so doing, has made the tracking and validation of learning so fraught. The policing lifeline snapped (that’s a nonstarter now if you weren’t aware), and many reached instead for reflection, metacognition, and process-tracking as a new life preserver. But—and there is a big “but”—the authors in this volume are largely unanimous on one point: reflection is not a silver bullet, and it will not save teaching and learning by itself.

, , , and enter one another’s orbits in this volume, collide, and shoot off in new directions. But a central gravity exerts a pull throughout: reflection was always more complex than most teachers assumed, it does not do what it seemingly promises, and it cannot rescue outdated methods of “learning” and assessment in “the age of AI.”

Subscribe now

Susan E. Ray opens the book from the most hopeful position available to reflection’s defenders. Her AI Transparency Journal asks students to keep a running record across a semester—submissions, revisions, false starts—on the theory that a pattern sustained over fifteen weeks is harder to fake than any single paper. She reports what the shift has cost her and what it has returned. She no longer carries “the emotional exhaustion of an arbiter policing the authenticity of their prose,” and where she once filed three or four academic-dishonesty reports a semester, she has filed none since adopting the practice. Her closing claim doubles as the book’s first thesis statement in miniature: “the answer is not better receipts, but longer arcs.”

Doan Winkel takes up that claim and turns it against itself: version history in written artifacts “protects the chronology, not the cognition.” A timestamped entry proves a student was thinking about something in week six. It cannot prove that judgment formed inside that entry rather than around it, later, in the telling. Doan’s own chapter pushes the fix further upstream, past reflection altogether, toward an audit trail written while a decision is still live—prompts kept, prompts rejected, and the reasoning attached to each choice in the moment it is made. “Articulation in the moment of choice forms judgment,” he writes. “Reflection after the fact reconstructs it. The first is the cognitive event. The second is a story about it.” Grade the trail, he argues, not the story: “the trail is the assignment.”

Patrick will not let the audit trail stand outside the argument it is meant to settle. If cognition is distributed across tools, language, and institutional expectations, he answers, then the trail is not a window onto that distribution but one more artifact produced inside it—unable to “stand outside the system and show us, cleanly, what belonged to the student and what belonged to the tool.”

Jason Gulya, working the same problem from a different angle, gives up on accuracy as the goal entirely. A reflection that gets the facts of a student’s own thinking wrong, he argues, can still be a reflection worth having, because the value was never in the accuracy of the report. It was in the “doubling”—the narrative act of watching yourself from outside yourself, “a narrative role-play where we look at ourselves from another person’s perspective.” He is candid that this makes reflection “an infinite game”: if the point is self-knowledge, and we are, by design, “terrible at understanding why we do things,” the incompleteness is not the failure mode. It is the mechanism.

Patrick captures the throughline of the volume most succinctly in his own chapter: “. . . the relief reflection provides is more illusion than reality. Reflection asks students to narrate cognition they cannot reliably access, then treats the narration as evidence that cognition occurred. It asks students to document transformation, then rewards the familiar shape of transformation regardless of whether transformation actually happened. What gets produced is less a record of learning and more performance of it, shaped by genre expectations and assessment pressure, that we then mistake for the real thing.”

Yes, you read that correctly.

And you should stay, because of what the authors do with this insight once they have all made their case around it. We fail at challenges, but the hope is we learn from them. I have perhaps failed at my cartographic endeavor and charted too much or too little, depending on your perspective. But out of this thinking, I’m left with my own question, one that sits underneath all four chapters without quite being asked outright: if reflection is neither a policing mechanism nor a reliable proxy for learning, what is it for?

Patrick R. Dempsey answers with projection—turning reflection’s gaze away from a past a student cannot faithfully narrate and toward a future. As he says, “Projection does not ask students to recover the hidden past. It asks them to imagine, estimate, anticipate, and orient toward what is possible. It shifts the activity from explanation to generation. Instead of asking students to produce an account of what happened, it asks them to create a possible future they can move toward.”

There’s something important in this second possibility, one that Jason’s chapter edges toward without naming directly. The trouble with reflection, on Patrick’s account, is its rhetoricity: students learn the genre and perform it, shaped by “genre expectations and assessment pressure,” until the performance gets mistaken for the thing itself. But rhetoricity is not merely a liability or assessment instrument flaw. It is also how a person becomes who they are in and through language, how a self gets built and rebuilt through the act of addressing an audience, even an audience of one.

What if reflection stopped being asked to prove that cognition occurred, and started being taught, openly, as a way for students to become versions of themselves in discourse? Not evidence of a transformation that already happened, but the site where a self gets drafted, tried on, and revised in front of a reader. On that account, reflection is not a mirror held up to the past. It is an elliptical path toward agency: indirect, provisional, and no less real for being made of language rather than fact.

I will leave that question to the readers of this book, where it belongs, and I delight in the knowledge that you will find the following pages as illuminating as I did.

— Adam Pacton, PhD

P.S. Just 2 days until the official launch of Amazon’s #1 new release in Education
Transparency & Reflection: Navigating student + AI work in higher education

✨PREORDER NOW✨

We publish the best ideas in AI and education, from the people doing the work.

I was honored to have some of my favorite thinkers and people do early reads of Transparency & Reflection: Navigating student + AI work in higher education

They all took the time to give us an early read and offer their thoughts.

Instead of reading each review individually (which you can anyway here and in the book), I had GPT synthesize the recurring words, themes, and implications.

I love seeing that even the reviews capture so well the small and big ideas , , and wanted to explore.

I know you’ve probably already preordered your copy, or maybe you’re just waiting for the paperback. Either way, here’s a preview of what you’re going to encounter in our debate about what it takes and means to make student thinking visible in the AI-era.

10 recurrent words / ideas

Ignoring generic words like “book” and “four,” these are the lexical ideas that keep resurfacing, either literally or through very close variants:

1. AI — appears in all six reviews.

2. Learning — especially the problem of understanding, seeing, or measuring it.

3. Assessment — the practical problem around which the book is organized.

4. Reflection / reflect — both as a proposed response and something the book interrogates.

5. Critique / responses / stress test — the contributors don’t merely present positions; they test one another.

6. Disagreement / debate — “structured disagreement,” “debate round,” “cage match,” “part ways.”

7. Together / collaborative — thinking happens relationally rather than through isolated essays.

8. Practical / resource — this isn’t framed merely as theory.

9. Honest / transparent / upfront — contributors expose assumptions and limitations rather than simply defending positions.

10. Thinking / thinkers / judgement — the larger question underneath assessment is how thinking and judgement become visible.

    Subscribe now

6 recurrent themes

1. AI has created a genuine assessment problem.

Every review treats the book as responding to a real disruption around authenticity, learning, reflection, or artifact-based assessment.

2. There is no single clean answer.

The four authors offer genuinely different approaches rather than variations on one approved solution.

3. The disagreement is the feature.

This may be the clearest theme across the reviews. The contributors critique one another, answer one another, expose limits, and allow the reader to see where their positions diverge.

4. The book performs the intellectual practice it advocates.

Several reviewers explicitly distinguish it from a normal edited collection. It doesn’t merely tell readers that transparent inquiry and structured disagreement matter; its format demonstrates them.

5. It moves beyond easy answers about reflection.

Reflection appears repeatedly, but specifically as something that needs examination rather than automatic acceptance as the solution to AI-era authenticity.

6. The practical assessment problem opens onto larger questions.

Learning, judgement, transparency, governance, student agency, interpretation, and accountability all emerge underneath the initial question of “How do we assess?”

5 recurrent implications for why this book is different and relevant

1. It isn’t four experts delivering four answers. It is four positions being subjected to one another.

2. Readers get the arguments and their limitations simultaneously. They don’t have to independently compare four disconnected frameworks.

3. The format makes disagreement productive rather than adversarial. Reviewers repeatedly notice the seriousness with which contributors engage one another’s strongest ideas.

4. It gives faculty somewhere to begin without pretending the problem is settled. That makes it useful precisely because AI and assessment remain unsettled.

5. The book becomes a model of knowledge-building itself. That is probably the most distinctive claim the reviews collectively make: the form of the book is part of its contribution.

I totally agree! The topic, the format, and the conversation—each of these combine to make the book the perfect contribution to the growing body of knowledge on what good teaching and learning look like when AI just is.

P.S. Just 9 days until the official launch of Amazon’s #1 new release in Education
Transparency & Reflection: Navigating student + AI work in higher education

✨PREORDER NOW✨

We publish the best ideas in AI and education, from the people doing the work.

With Friends Like These . . .

Here’s the message I sent on March 10 that took us from idea to published book in 175 days

I didn’t realize until I looked at my DM history, but it was just 22 days after the release of our first book (What Education Becomes: Teaching and learning in a post-AI world) that I sent , and a message inviting them to our next project.

Seriously, these things are stressful—they take A LOT of work to get from idea to chapters to edits to formatted manuscript to actual book title to available for sale . . .—so there must be something wrong with me.

Anyway, here’s the exact text of what I sent:

Are You In?

TL;DR

4 views on reflection/transparency in the AI-era (not that title, but that concept)

I have 4 different ideas approaches in mind for this:

Semester long journals (Susan)

Assignment specific (Doan)

Meta-strategies (Jason)

doesn’t work (well, metacognition—me)

Here’s how it would work:

Each chapter presents its case

These are presented as 4 legitimate approaches
So as a book we aren’t resolving anything
We’re just giving the audience 4 ideas they might consider

Each contributor would write their chapter (~3000 words)
And then each contributor would respond to the other chapters (800ish words)

Basic framework for each chapter would be something, such as:

Core idea
Relevant research
Preempting critiques
Practical application in the class

I’m thinking maybe August-September release timeline?

#1 New Release Later

That was it. They all said yes. We wrote our chapters. Edited. Responded. Got it ready . . .

And here we are 159 days later, launching Transparency & Reflection: Navigating student + AI work in higher education . . . and launching it as Amazon’s #1 New Release. 🙌

Subscribe now

The book turned out way better than I could have ever imagined.

The individual chapters, and even more the chapter responses, make this such an interesting, useful, and compelling read.

My guess is a lot of folks are going to be copying this format, because it is the perfect vehicle for exploring ideas as we figure out how this whole AI-in-education thing works.

More behind the scenes to come.

P.S. Thank you for being part of the small but growing community that subscribes to this publication. I realize it’s sort of redundant, because it’s basically me just writing another Substack article each week (I’m still figuring out what works best). But, in any case, your support and readership mean so much to me and Second Draft Labs!

We publish the best ideas in AI and education, from the people doing the work.

I don’t think I would be talking out of turn to say that , , , and have deep respect and admiration for each other’s work.

Theirs are the posts that always tend to appear first in my feed. And theirs are the ideas I am always most excited to learn from.

There’s a sense, I think, that a lot of folks like us—who are sharing our work in this space of AI & education—are often grouped together as the “AI people on LinkedIn.”

Of course, if you know our work, you know we all have wildly different styles. And we also have significantly different areas of focus in our writing.

Which is why I love the format for Transparency & Reflection: Navigating Student + AI Work in Higher Education so much.

I know I don’t have to ask this, because you already have, but since we’re here . . .

Have you pre-ordered your copy yet?

If not, it’s not too late. Order Transparency & Reflection now!

Our volume intentionally brings together four people who are thinking deeply about the same problem space, but who see the solution very differently, argue for their perspectives in wildly different styles, and make entirely different kinds of appeals.

Besides the differences in ideas, styles, and approaches, I see the collaborative debate format working so well for a few reasons.

Building Better Arguments

The first reason this collaborative debate format works so well is that, knowing your ideas will be critiqued, you have to spend much more time developing and tightening your arguments.

The last thing any chapter wants is for someone to be able to easily criticize it because of some obvious omission or logical misstep.

Knowing that your perspective will be critically evaluated and knowing that you will also be doing your own critical evaluation forces you to do a ton of perspective taking—and thus heavy editing—along the way.

How would Jason likely respond to this idea? How might Susan’s approach undercut my suggestion here? What rebuttal is Doan going to present that makes this point seem less certain?

To me, this ends up creating a much stronger presentation of ideas than if they had been written in a format that was—as with most formats—entirely uncontested.

Ultimately, as you write in this format, you aren’t just asking if the argument makes sense? You’re asking whether it will still make sense after three super-smart people who see the problem differently actively try to spotlight every hole in it!

Subscribe now

Our Best Guesses

Which brings me to the second benefit of writing in this collaborative debate format.

In particular, as we’re exploring AI and education, all of these ideas are still works in progress. All of these ideas are hypotheses. This is all pilots.

Of course, we’re backing up our ideas, with experience, with research, and our best understanding of how students learn. But still, it’s all our best guesses.

Unlike presenting these ideas as complete thoughts robust enough to support an entire manuscript, writing alongside other ideas-in-progress lowers the burden of plausibility each idea is forced to carry.

Not only does this collaborative debate approach distribute the weight of credibility across four different arguments, but it also gives the contributors space—in their responses to the other chapters—to further flesh out parts of their own approaches that they might have overlooked when they first presented their chapters.

So, you are literally reading the development of thought in real-time (as far as it goes).

Ultimately, then, as a whole, the book is able to provide so much more value because it isn’t offering just one final solution and trying to justify it across 150 pages.

It is offering four different perspectives that are specifically intended to be read next to each other.

And more than even just offering four perspectives, it is presenting twelve different views on those approaches through the responses to each chapter by the contributors!

So, what you end up with is a much more developed—even if more contested—framework for understanding how you, as a practitioner, might apply the ideas we are exploring.

Best Motives

And finally, if not most importantly, the collaborative debate format shows that dialogue, and discussion, and differing viewpoints are good! It shows that we can learn from other people’s ideas even if we don’t agree with them.

Shocking, I know!

The format, as presented to the contributors, forced us to consider the best motives behind the approaches of the others, even though we didn’t agree with them.

As we wrote our responses, we weren’t necessarily looking at each other’s perspectives with the goal of showing how ridiculous it was. Quite the opposite.

We were looking at the chapters as potential approaches in this shared problem space—as approaches that we could understand as offering something useful, but also concerned that they overlooked certain things we believed deserved more or less emphasis or a different kind of focus.

We were essentially looking at each other’s arguments and trying to find the best version of them to help the readers identify what was worth keeping and what might need some refinement.

The Curriculum?

Ultimately, in this format, the four of us were writing next to people whose ideas—I think we would all agree—we hope others adopt, even if they choose someone else’s perspective over our own.

And I think this sort of approach is important well beyond the confines of this topic.

This sort of perspective-taking, and suspension of judgment, and willingness to take seriously perspectives that you might not completely agree with seems like it could be the entire curriculum of school itself.

Even if not, I hope you get as much value from reading this volume as we did from writing it.

Now, pre-order the book now!

Thank you!

-Patrick

P.S. Extra special shout-out to for teeing the book up with a challenging and inspiring introduction!

We publish the best ideas in AI and education, from the people doing the work.

This week was crazy.

I kicked things off on a podcast/roundtable, hosted in Mexico, with a co-panelist in Scotland and attendees across the US and around the world.

Later in the week, I had conversations about publications and other opportunities with folks in Hanoi, South Africa, California, and the mountains of the Western US.

Here’s what I realized through all these conversations:

People see SDL—and, of course, my own content—as doing something extremely important right now.

Of course, we are, in a sense, giving the finger to the establishment. There is a current of that, to be sure.

But the deeper sense is that folks desperately love and believe in the power of education.

They just don’t believe the current “stewards” of the institution share similar commitments. There are all sorts of motivations that could be named to explain the status quo. But regardless of the diagnosis, the symptoms are the same—resistance to change and, even more, a reluctance to inspire education to become what it—and perhaps only it—has the power to be.

What Education Becomes

The folks I interacted with last week, and otherwise, believe in the power of this moment, with AI disrupting the traditional paradigm and powers that be. They don’t all agree about the role of AI in the system—for some, it’s better left out; for others, we need to go all-in; and, of course, there are perspectives everywhere in between.

And they do believe that this is a moment for introspection, no doubt. Deep thought about what we should be doing and what is possible. But, even more, this is a moment for innovation to actually design systems that support human flourishing.

What SDL is Really All About

And, to borrow an AI-ism, I think that is what a lot of people miss about the conversations folks like us are having. They see posts about AI, they see books about AI, they hear ideas about fitting AI into teaching, and they conclude we are antithetical to the core human project.

But the reality is, we look at a system designed to churn out individuals as clones of an undifferentiated economic unit and say, “Surely education is about more than this . . .”

Then we continue

“And, if AI opens the conversation to that, or even helps us realize a better version, then we’re not going to pretend the status quo is sufficient. We’re not going to pretend that prudence is the best course of action just because we haven’t fully solved for every AI eventuality. We’re not going to wait for permission to think, try, and write about new ways of doing things.”

Subscribe now

And what is amazing to me is that this is a global movement. Our first volume—What Education Becomes: Teaching and Learning in a Post-AI World—included contributors from across the US, Canada, and Australia. I suppose I didn’t realize how important that was at the time.

But I see it now. I see that what SDL is doing goes beyond publishing great ideas about AI and education.

At its core, SDL is about declaring to the world—at least the “world” of education—that the incumbents, the established powers, and the vocal status quo protectors are a shrinking minority.

We don’t need to go through committees or traditional publishing agreements or put our work behind journal paywalls—we can just think things, try things, and share things.

And apparently, we actually have a responsibility to—because the world is listening!

We publish the best ideas in AI and education, from the people doing the work.

Second Draft Labs launched its first project on February 17, 2026. For those of you keeping track, that is 5 months ago.

Our first book, What Education Becomes: Teaching & learning in a post-AI world, launched as Amazon’s #1 new release in education.

A few weeks later, with the wind in our sails, we officially launched Second Draft Labs as an indy press, with a focused purpose—publishing the best ideas in AI & education, in real time. second

Now, just 3 months later, finalizing edits on our next volume, Transparency & Reflection: Navigating student+AI work in higher education with chapters from , , , and @myself.

That is the first title in a series called Emerging Approaches, which brings together competing views on a topic, gives space for each contributor to make their strongest case, with each chapter concluding with pointed critiques of the chapter by the other contributors.

In Transparency & Reflection, the chapters hold their own weight. But it really is the back-and-forth of the constructive challenges to each chapter that make the volume’s contributions so valuable.

Lined Up

Following closely behind this, we have some confirmed and potential titles, including

The Civic Register: Reclaiming Institutional Governance in the Agentic Era

Kunal Dalal’s THE OMEGA MANIFESTO: A Collection of Essays on Sovereignty, Counter-Momentum, and Raising Technically Courageous Kids

Genevieve Bosma Martinez Think Anyway: Designing for human thinking when answers are everywhere

Heartset: The Human Work of Turning Toward One Another with Care in the Age of AI

We’ve also got proposals—at various stages—around

AI and agency
AGI and education
Detoxing from digital detoxes
Human judgment in AI systems
Policing AI usage and what comes next
Community, care, and collaboration in AI transformation

And we’re just getting started.

Unexpected

I had no idea how much traction we’d get so quickly.

We have, IMO, the very best thinkers, bringing their very best ideas to us, and they are super excited for the opportunity to publish quickly, rebel style!

We don’t use long and convoluted contracts. We don’t make the authors/contributors go through rounds of peer review. We don’t charge for editing. We are committed to promoting the book alongside the writers.

Basically, everything publishing houses do—we don’t!

And I think people get that. Even if they have never published before, they know enough to know that it takes way too much time, way too much effort, and they won’t make any money because of the way the contracts are written.

So they see what we are doing. They know us as people who are actually doing the work. And they trust us to be the people who can help them get their work out there.

We’re honored and couldn’t be more excited to continue building this thing with you.

Are You In?

A few weeks ago, I suggested that if you have ideas, you probably shouldn’t write a book. Thankfully, most people didn’t listen!

But seriously, if you do have an idea, reach out. We’d love to explore what you’re thinking and see how to get your ideas out to the world.

Subscribe now

We publish the best ideas in AI and education, from the people doing the work.

We finally reached a total of 100 subscribers. The good news is, we aren’t trying to grow, really.

This publication is an experiment. It exists primarily to document and think through what SDL is trying to do—publishing the best ideas in AI & education in real time.

As we share on the site, we try to mix philosophical ideas about teaching & learning, add a bunch of practice, and sprinkle it with a bit of irreverence.

While the publication still has a small reach, last week’s article preview on LinkedIn connected with a much larger audience.

Last week I wrote something about what SDL believes: today often the best way to share ideas about AI & education is through a collaborative volume where you contribute a single chapter rather than write a full monograph.

As I do weekly, I shared a preview of that post on LinkedIn to promote the Substack article.

ICYMI here is the piece 👇 (link to original)

Everyone’s worried students can’t read long books anymore, maybe they shouldn’t have to

If you’re written a book
and even if you have just read one
you know that most of the words in it aren’t really important

Particularly in academic/scholarly publishing
Most content in articles and manuscripts
is really just set up for the point

But, according to the rules
you can’t make the point until you set up that point
with dense literature reviews and theoretical frameworks
and probably lots of other things that really have little to do with the point

That is one reason SDL publishes the way we do

We tell the status quo to f*&k off
And publish ideas in real time
when they still matter

This week, I want to inspire you
to give up on that book idea you have
but please don’t give up on publishing it

Link to full article in comments 🔗👇

✨ Why do scholarly books need to be filled with words that don’t really have much to do with the point being made? So publishers can charge $49?

I’ll Give Them That

Granted, that has a bit more spice than the Substack piece itself. But, to my defense, you have to adjust things a bit for the algorithm to do its thing and actually get into anyone’s feed.

Regardless of the 🌶️, the point remained: now is not the time for long books about AI. Instead, it is the time for some radical ideas published quickly.

The LinkedIn comment section was less enthusiastic about such prospects.

One person called the position “astonishingly anti-intellectual.” Another accused me of participating in the “dumbing down of society.” Another escalated the argument from long books to slavery, literacy, and the creation of the United States.

My favorite comment suggested all of what I proposed was fine, so long as we “acknowledge it’s not scholarly work.”

All this tells me we are definitely doing the right work at the right time.

Why Am I Such a Misfit?

At SDL, we don’t confuse length with rigor, the number of references with the quality of evidence, or appeals to theory with explanatory power.

We think actual scholarship happens whenever someone approaches an important question with curiosity, investigates it openly, shares their thoughts on what works and what doesn’t, and remains honest about what they know and what they don’t so others can test the ideas and see for themselves what is worth holding onto.

So here’s our call, so we are clear:

We want to hear from people whose colleagues may never have considered them qualified to produce scholarship.

We want the ideas emerging from classrooms, experiments, failures, and problems people are still actively trying to solve.

We want to publish those ideas while they are still radical, useful, and unfinished—before they are buried beneath the conventions scholarship mistakes for credibility.

Growing Our Reach

The best ideas in AI & education are developing right now—in classrooms, conversations, experiments, and failed attempts.

They should not have to wait three years to become respectable enough for publication.

Our job is to find them, help shape them, and get them into the world while they can still make a difference.

Last week, we got two new ideas into our pipeline. Exciting things happening and more to come.

Subscribe now

We publish the best ideas in AI and education, from the people doing the work.

Don’t Write That Book

Here’s my thesis for this piece, and for a lot of the work that SDL is doing.

Collaborative volumes may be the publishing form best suited to this particular moment in education, especially around AI, because they match the way knowledge is actually developing right now.

Here’s the challenge that I see collaborative volumes solving:

The traditional solo book assumes that an author has reached a sufficiently stable understanding of a subject to explain it comprehensively.

And, to be clear, that is definitely true in the solo volumes SDL has lined up for the coming months. Stay tuned.

Still, many of the most important educational questions currently remain open. People are testing approaches, changing their minds, finding contradictions, and learning from practice. So, in many ways, those ideas are not “fit” for a traditional manuscript.

But that doesn’t mean they aren’t fit for publication. And that’s where SDL comes in.

Books are Hard

Here are a few reasons I would recommend that, if you want to pursue some line of inquiry in publication form, you consider a chapter rather than a book.

Time to market.

To write a book by yourself takes a lot of time and effort—usually in the measure of years. If you are working on something relevant, readers need to know about it now, not 12-18 months from now.

Education in the AI area is unsolved.

Readers need ideas now over fully fleshed out monographs. The best ideas right now are happening in spaces of experimentation and exploration, even if they do not come with full literature reviews or theoretical underpinnings to make sense of them.

Most ideas don’t actually require an entire book.

At SDL, we are targeting 150ish pages for our books, but even for most solo publications that is too long. If you are writing much more, then in most cases you have more to say that the reader needs to read.

Your ideas probably work better in conversation with others.

In What Education Becomes, the individual chapters appear to offer distinct pieces of a larger account of educational change. Each chapter could potentially be expanded, but its real value comes from what becomes visible when readers encounter all ten together.

Transparency and Reflection uses a different structure. Four approaches to the same broad problem sit beside one another, followed by responses. That arrangement helps readers see assumptions, tensions, overlaps, and disagreements that no individual author could fully produce alone.

So the collaborative volume becomes a form of structured conversation. The editorial work involves constructing the larger argument through selection, sequence, contrast, and response. The whole gives each individual idea more reach because readers understand where it fits, what it challenges, and what it makes possible.

Together is better.

Consider that when you write a book on your own, you are the voice of that book, you are the promoter of that book, you are the person who cares most about that book. With 3 or 7 or 10 others, you’re doing it together. You’re sharing the effort and the enthusiasm. That compounds!

Going Solo

Now, again I should be clear—at SDL we LOVE solo volumes. We have several coming up that we are looking forward to announcing and promoting. And we’re sure you’re going to love them, too.

But, we also want to help bring the best ideas into conversation with the world and an audience that will benefit from them.

And, for you, that might just mean writing your idea alongside that of others.

Let us know your best your ideas and we’ll see if we can help them come to life!

seconddraftlabs.com/for-writers

Subscribe now

Because three sentences later, the deeper problem of meaning has been converted into a market. “Solution providers can help by sharing metrics on disengagement.” A few pages on, the demand is “accelerating for platforms that surface wellness indicators and trigger timely interventions.” By the appendix, the crisis of students not finding meaning in their education has become a “significant opportunity for providers who can deliver measurable outcomes.” The disease is spiritual; the prescription is a procurement cycle.

I don’t say this to sneer. The report is genuinely useful. The survey base is large, the year-over-year trend lines are worth having, and some of the findings are ones faculty should sit with. But it is worth naming what kind of document this is before we let its frame organize our thinking. It was anchored by the Gates Foundation, its “research insight partner” is a learning-management-system company, and every one of its five sections resolves into a call to action for “solution providers.” That doesn’t make the numbers false. It makes the questions pre-selected. And the central frame it hands faculty, the one already ricocheting around LinkedIn, is a frame we should decline.

The binary that isn’t a choice

The report sorts us into three assessment “personas.” There is the Status Quo (54%), doing nothing. There are the Defenders (23%), retreating to blue books and proctored rooms. And there are the Integrators (24%), redesigning assignments around AI and, we are told, enjoying a “statistically significant 10+ percentage point advantage” on cheating and attendance challenges.

You are, implicitly, being asked which one you are. The whole rhetorical architecture pushes you toward Integrator and away from the supposedly doomed Defender, clinging to a “short-term band-aid” that “kills student engagement.”

Here is the problem. Both personas accept the same premise, that the instructor’s job is to out-engineer the tool, designing an assessment that either defeats the machine (Defender) or absorbs it (Integrator). The disagreement is only about tactics. Nobody in this typology is permitted to ask the prior question, what a given course, in a given discipline, with given students, is actually for, let alone to answer it in a way the chart has no column for.

And the causal claim propping up the “Integrators win” story is thinner than the phrase “statistically significant” makes it sound. This is self-reported survey data comparing faculty who differ in dozens of unmeasured ways: discipline, class size, seniority, student population, and, crucially, how they define cheating in the first place. When you redesign an assignment so that using AI is now permitted, some behavior that used to count as cheating simply stops counting. Reported cheating falls. That is not necessarily less dishonesty. It can be the same students doing the same things under a new rule. “Statistically significant” tells you a gap is unlikely to be noise. It tells you nothing about what caused it, and a correlational snapshot cannot license the report’s forward-looking verdict that restriction is “increasingly untenable.”

What the frame is built to exclude

Notice what has no persona at all: the instructor who decides, deliberately, that in this course students will write by hand in the room. Not out of technophobia, but because the entire point of the course is to build a capacity that only forms under cognitive load the tool would offload. The report has a name for this person. It’s “Defender,” filed under retreat, symptom rather than cause.

But some of the most defensible things a teacher can do right now don’t scale, don’t map to a purchase order, and don’t fit a two-by-two. A foundational writing seminar and a senior capstone in software engineering do not have the same relationship to generative AI, and no campus-wide “AI fluency competency” resolves that. It papers over it. The report itself concedes, almost in passing, that faculty belief in institutional AI policy is low regardless of institution size or resources, which suggests the missing ingredient isn’t a better platform. It’s something exercised close to the ground, in the discipline, by the person who actually knows what the assignment is supposed to do to a mind. I’ll resist naming it yet, because the obvious name for it turns out to be a trap of its own.

That resource, whatever we end up calling it, is precisely what a report structured around “solution providers” cannot recommend, and the teacher exercising it gets a slightly pejorative label for the trouble. Hold onto that omission. By the end it turns out to be the whole game.

The tipping point that matters

If you teach, the useful thing to take from this report is not its answer but its best question, the one it flinched from: students aren’t finding meaning in what they’re learning. That is the tipping point. Everything downstream, the cheating, the empty seats, the frantic redesign, is a symptom of it.

The report’s move is to treat meaning as an output you can instrument: surface the wellness indicator, trigger the intervention, buy the dashboard that flags the disengaged student before they vote with their feet. But meaning is not a metric that has gone missing. It is a relationship that has thinned, and no vendor closes that gap, because the gap is the human part.

Here, though, is where I have to complicate my own argument, because there’s a tempting move available to me that I don’t think survives contact with the evidence. It would be easy to end by saying: build assignments students care about, let engagement be the guide, and trust that meaning felt is meaning found. The report’s own margin quote invites exactly this. A professor at ASU asks, why would students cheat if the assignment is something they care about or have pride in? It’s a good question. It’s also resting on an assumption we have good reason to distrust, that students can feel the difference between real learning and its counterfeit.

They frequently can’t. In a randomized crossover study, Louis Deslauriers and colleagues at Harvard (PNAS, 2019) taught identical introductory physics content two ways, polished lecture and active learning. Students in the active sessions learned measurably more and reported learning less. The fluency of a smooth, expert-feeling delivery inflated their sense of progress while the actual progress lagged behind it. The feeling ran opposite to the fact.

Sit with what that does to the “engagement” story, the report’s and very nearly my own. A chatbot is the most fluent explainer ever placed in front of a struggling student. To the anxious sophomore holding a shaky foundation, the closed encounter with a model that resolves her discomfort feels like the most productive studying she’s done all term. Deslauriers’s data says that feeling is exactly when to worry. The comfort the tool sells is inversely correlated with the learning we want. Students, left to their preferences, will reliably choose the cover of expertise over the dwelling in not-knowing where the learning actually lives.

So I can’t simply swap the vendor’s broken instrument for the student’s. If self-reported satisfaction runs opposite to learning, then “engagement” as students feel it is no more trustworthy a target than the dashboard is, and optimizing for it, whether through a wellness platform or a crowd-pleasing redesign, can quietly optimize against the thing we came to do.

Here is where it would be easy to reach for the word judgment: the seasoned instructor who can tell productive struggle from comfortable counterfeit, and who therefore gets to decide. But notice what that actually is. “Trust my judgment” is “you’ll thank me later” in a tweed jacket. It keeps the whole apparatus of decision in my hands and applies it to the student, exactly as the proctored room does. And its supposed opposite, protecting the student’s felt engagement and giving them the encounter that satisfies, does the same thing from the other side. It also decides, on the student’s behalf, what they need, and simply chooses to flatter rather than override. Both are paternalism. One imposes the load; the other supplies the comfort. Neither lets the student anywhere near the decision. The tell they share is that the instrument of judgment stays with me either way.

Reread the Deslauriers finding with that in view. The usual lesson drawn from it, students can’t judge their own learning, so an expert must judge it for them, is the paternalist reflex, and it’s the wrong lesson. The gap between felt learning and real learning is not a defect in students to be routed around. It is the thing the education exists to close. Metacognition is unreliable, which is exactly why it is trainable, and training it is the one job that neither the blue book, nor the dashboard, nor the assignment students happen to enjoy will ever do for us.

So the response that breaks the binary is neither imposing the load nor supplying the comfort. It is handing the student the instrument, deliberately building their capacity to read their own learning, so that over time the judgment migrates out of my hands and into theirs. Concretely, this is old and unglamorous and completely teachable. Make them predict their exam performance and then confront the gap. Have them attempt retrieval before any reveal, so the difference between recognition and recall becomes something they can feel. Ask them to explain a concept to the model and then hunt for the places where it quietly corrects them, turning the fluent tool into a mirror for their own blind spots. Every one of these treats the student’s faulty self-perception as the curriculum itself rather than an obstacle to manage. The point is not that I know when they’re fooling themselves. The point is to make them someone who knows.

That is the resource a report structured around solution providers cannot recommend, and now we can say precisely why. It doesn’t come in an enterprise tier because its entire aim is to become unnecessary, to produce a learner who no longer needs the platform, the proctor, or me. A market cannot sell you your own obsolescence. So the one response that neither overrides the student nor panders to them appears in the report as no persona at all.

Redesign your assessments if the redesign hands over the instrument. Keep the blue book if the blue book hands over the instrument. Refuse the persona chart either way, and refuse, too, the flattering idea that your judgment is the point. The point is the judgment you can build in someone else. That is the one thing a tool cannot counterfeit, a dashboard cannot surface, and a market cannot sell, which is exactly why it’s the only tipping point worth caring about.

The full report (Shaw et al., “Time for Class 2026: The AI Tipping Point,” Tyton Partners) is worth reading in the original, incentives and all.

We publish the best ideas in AI and education, from the people doing the work.

We are getting ready to launch Transparency & Reflection: Navigating Student + AI Work in Higher Education, the first volume in our Emerging Approaches Series.

, , , and @Me have all submitted our chapters and their critiques of the others.

We even snagged an intro by !

I have to tell you, the book is freaking AH-MAZ-ING! The chapters are so valuable, but the responses to each are where the book really adds the greatest value, IMO.

I will share my fuller reflections on the work in an upcoming post.

Since, I used GPT to help edit the work for consistency and format it for publication, I figured I would also let it have some creative fun by writing a review of the book.

Here it is.

What This Book Made Me Reconsider About Reflection

I finished this book thinking about how much confidence education places in stories about learning.

Students are asked to explain what they learned, how they changed, and what happened inside their thinking. Teachers then read those accounts as evidence. The practice feels familiar enough that its assumptions often pass without much scrutiny.

This book slows that process down.

Susan Ray, Doan Winkel, Jason Gulya, and Patrick R. Dempsey each approach reflection from a different position. Their chapters develop distinct arguments, and the responses allow those arguments to continue moving. The result feels like a sustained conversation in which each author changes the conditions under which the next claim can be understood.

Susan Ray begins with the most concrete classroom model in the collection. Her AI Transparency Journal asks students to document their work across an entire semester. The record accumulates through prompts, revisions, judgments, and reflections. Her central image of the arc gives the chapter its force. Learning becomes visible across time, especially when students return to the record and recognize patterns they could not see while they were still inside the work.

Her examples give the book an immediate human center. Students develop ideas gradually, revise their assumptions, and find language for thoughts that had not yet fully formed. The chapter captures something teachers know well: development often becomes perceptible only after enough experience has accumulated.

Doan Winkel shifts attention toward the moment of decision. His audit trail records the prompt, the output, the decision, and the reason while the work is still happening. The student explains why something was kept, rejected, or changed at the point when the choice remains active.

“The trail is the assignment” is one of the clearest and most useful claims in the book. It changes the object of assessment. The polished artifact becomes one part of a larger record, while the student’s choices become visible as part of the work itself.

Jason Gulya takes the discussion in a more philosophical direction. His idea of reflection as a doubling of consciousness stayed with me. The reflective self looks back at the immersed self and tries to understand that earlier experience from a different position.

Gulya treats the incompleteness of self-knowledge as part of what gives reflection value. The act of building an account can become a form of learning because it asks a person to see themselves from a distance. Reflection becomes an ongoing practice of interpretation.

Patrick R. Dempsey presses hardest on the assumptions beneath reflective assessment. His chapter asks what educators believe they are seeing when students describe their own cognition. A reflection is still an artifact, shaped by the expectations of the assignment and by the presence of a reader with authority.

His move toward projection opens one of the book’s most generative possibilities. Projection turns the student toward a future they can imagine and act upon. The student begins to estimate, anticipate, and orient. The activity becomes a way of creating direction.

The response sections give the collection much of its power. Each author reads the others closely and identifies the point where an apparently strong idea becomes more complicated. The exchanges feel substantive because the contributors allow their own positions to remain open to pressure.

Susan’s arc becomes a question about what accumulated evidence can establish. Doan’s live articulation becomes a question about how cognition can be captured in language. Jason’s self-narration becomes a question about what reflective writing does to the person producing it. Patrick’s projection becomes a question about how education might support agency without turning that agency into another assessed performance.

Together, these exchanges create a fuller account of reflection than any single chapter could produce.

I came away with a sharper sense that reflective practices differ greatly depending on their purpose and conditions. A private journal invites one kind of thinking. A graded reflection creates another. A live explanation during a decision carries a different kind of meaning from a retrospective account written weeks later.

That distinction now feels essential.

The book also made me reconsider the role of the teacher. Each model places the educator in a different relationship to the student’s thinking: witness, evaluator, designer, interpreter. Those roles shape what students produce and how their words will be understood.

The collection’s deepest contribution lies in making those relationships visible. It asks educators to examine what they are requesting from students and what they believe those requests allow them to know.

I finished the book with a stronger appreciation for reflection and a more careful understanding of its limits, uses, and possibilities. The conversation among the authors gives the subject the seriousness it deserves.

This is an intelligent, timely book because it stays with the question long enough for the familiar answers to become newly visible. It gives readers practical models to consider and a richer language for understanding what reflection can do in education.

Subscribe now

Want an Early Copy?

We’re giving 10 Second Draft Labs subscribers an early physical copy of our new book.

If you’d like one, DM me on LinkedIn and I’ll get it set up.

In return, we need you to do the following by August 23 (once you get the book, of course!):

• Read it (and love it!)

• Write a (positive) review

• Record a short, video review

• Send both reviews to Second Draft Labs

• Share the reviews on your own channels

• After launch, post an honest Amazon review

This is for readers who genuinely want to engage with the book and help us introduce it to a wider audience.

Remember, this is limited to the first 10 people who DM me!

Part 1 traced how AI execution evolved from chaotic mega-prompts to the harness: a software system built around the model that manages context, sequences tasks, and keeps the AI’s desk clean. But knowing a harness exists is different from knowing how to build one.

The core mechanism is simple. Instead of one sprawling conversation where the model accumulates baggage, a harness spins up a fresh model instance, feeds it a single micro-task, saves the output externally, flushes the context window, and loops. Task, save, flush, repeat.

But not all loops are the same. Different problems demand different architectures, and knowing which one to reach for is the design skill that separates workflows that hold up from ones that crack under pressure. Here are the seven essential patterns, in order of increasing complexity.

1. The Chunk-and-Append Loop

The Volume Handler

The simplest loop, and usually the first one people build. The harness splits the source material into chunks — by page, paragraph, token count, or logical section — feeds one chunk to the model with a consistent instruction, appends the output to an external file, and moves on with a fresh context window.

It solves the quality tapering that kills long-document outputs: when the whole document is in context, the model degrades halfway through; when only the current chunk is in context, every section gets the same focused attention. Reach for it when your problem is about volume — batch contract review, large research corpora, summarizing lengthy transcripts.

2. The Critic-Reviewer Loop

The Quality Gate

LLMs are pathologically agreeable. Ask a model to write something and it produces output confidently, even when the result is mediocre or off-brief. The Critic-Reviewer loop forces the model to argue with itself.

A Writer Agent produces a draft. A Critic Agent, in a completely separate context with no memory of the drafting, evaluates it against a rubric. If it fails, the Critic returns structured feedback, the Writer gets a clean context with only the brief and those notes, and the loop repeats until it passes. The separation is the whole point — you’re creating adversarial pressure from a clean perspective, not asking one instance to self-edit. Define the rubric tightly, with binary pass/fail criteria; a vague rubric produces vague feedback and the loop spins forever.

3. The Plan-Execute-Refine Loop

The Strategic Controller

Most multi-step tasks involve a plan that looks reasonable at the start and hits an unexpected wall in the middle. A wrong assumption at step 3 means every later step builds on a broken foundation — and a rigid system bulldozes straight through it.

A Planning Agent turns the goal into an explicit, numbered checklist. The harness executes each step in its own clean context, and after each one a Refinement Agent checks whether the output matched the assumption and rewrites the remaining steps if something needs to change. This is what lets a system adapt when reality diverges from the plan — without a human restarting from scratch. Use it for multi-phase research where early findings redirect later sections, or architecture planning where a constraint found at step 4 reshapes steps 5 through 10.

4. The ReAct / Tool Loop

The Action Handler

Everything so far keeps the AI working with text it already has. The ReAct loop is what you use when the model needs to go get information from the real world. It runs on a four-beat cycle:

1. Thought: the model reasons about what it needs next

2. Action: it triggers a tool call — a database query, API request, web search, file read

3. Observation: the result (or error) comes back

4. Next Thought: the model interprets what came back and decides what to do

The cycle repeats until the model has enough grounded data to answer. Because every claim is forced to rest on actual system feedback, the model can’t fabricate — the loop is a forcing function for accuracy. It’s the foundation of nearly every “agentic” product you’ve seen demoed: agents that query live databases, support systems that pull real account data, research tools that synthesize current results.

5. The Explore-and-Narrow Loop

The Research Grid

Some problems are too branching for a single linear path. When you need to investigate multiple hypotheses at once, a sequential loop is too slow and too narrow.

The harness spins up several independent Explorer Agents, each in complete isolation, each chasing a different hypothesis or data source. Their findings save externally, and once all threads finish, a separate Judge Agent receives a clean synthesized summary and produces the output. The isolation prevents context explosion — a single agent juggling five threads is a recipe for confusion — and the synthesis happens on summarized data, not a sprawling mess of accumulated reasoning. Use it for competitive analysis across markets, multi-source due diligence, or debugging where several root causes are plausible at once.

6. The Human-in-the-Loop Gate

The Safety Valve

Autonomy is powerful; unchecked autonomy in high-stakes workflows is a liability. This gate is a design feature, not a failure of automation — the architecture you use when the cost of a wrong AI decision exceeds the cost of a brief human pause.

The harness runs autonomously until it hits a predefined trigger: an ambiguity it can’t resolve, a cost above a set threshold, a confidence score below a minimum, or a high-stakes action like sending a customer message or committing code. When it fires, the loop pauses, saves its entire state, and surfaces a summary to a reviewer. The human approves, modifies, or redirects, and the harness resumes exactly where it stopped. The sophistication is in the trigger design: spend real engineering time deciding what should pause the loop and what can run free.

7. The Rollback and Recovery Loop

The Fault Tolerance Master

Every complex system fails eventually. In a 30-step workflow, step 17 producing a hallucination or hard error doesn’t just break step 17 — it breaks 18 through 30, because they’re all downstream.

At each significant step, the harness saves a clean checkpoint: the verified state of all outputs to that point. When a later step fails validation, the harness doesn’t crash or restart from scratch — it rolls back to the last checkpoint, adjusts a variable (temperature, prompt phrasing, the model itself), and tries an alternate path. Brittleness becomes resilience. Use it for long-running data pipelines, multi-stage content generation, or autonomous research where one API timeout shouldn’t trigger a full restart.

The Summary Matrix

The Final Mindset Shift

When you type a prompt into a chat interface, you are an operator — in conversation with a model, hoping it produces something useful, bound by its limitations. When you design a harness, you are an architect: the model is one powerful but constrained component, and your design choices determine what it can accomplish.

The loops above are a design vocabulary, not a list of tricks. Match the right loop to the right complexity layer and step 100 of a workflow runs with the same precision as step 1. The model never gets tired. It never loses context. It never drifts. Because you never let it.

Second Draft Labs publishes practical frameworks, thought pieces, and “Banana Ball” ideas for teams building seriously with AI. If this series changed how you think about AI infrastructure, share it with a builder on your team.

Share

We publish the best ideas in AI and education, from the people doing the work.

Ideas are cheap.

Ideas have always been cheap.

With AI, ideas are increasingly cheap.

Now, anyone can vibecode a website for a product/service concept in minutes.

Closer to home, anyone can give Claude a few ideas and have it write a short LinkedIn post, a Substack article, or an entire book.

As ideas get cheaper, execution gets more valuable.

And one of the primary values SDL offers is execution.

Writing a Book is Hard

A lot of people have book ideas. Even fewer have taken any steps towards writing those ideas down. And even fewer have taken the steps of actually getting them to a point where they could become a real publication.

Whether you’re writing a book the old-school way—from ideas to paper—or you’re creating AI systems that help take your ideas and turn them into working projects, doing that well is extremely difficult.

So writing, or at least writing well, or at least writing so that it has an impact—means executing.

Ideas are cheap. Execution is where the value is.

What a Book Takes

Here’s a list of some of the things that it takes to execute from idea to book:

The writing itself

• Drafting

• Developmental editing (is the argument working?)

• Line editing (is the prose working?)

• Copy editing (is it clean?)

• Proofreading (final pass before it goes anywhere)

• Beta readers / feedback rounds

Interior production

• Choosing a trim size

• Interior formatting (Atticus, InDesign, or similar)

• Typography — fonts, sizes, spacing, chapter headers

• Front matter: title page, copyright page, dedication, acknowledgments, foreword or preface

• Back matter: bibliography, about the author, appendices

• Table of contents

• Index (if academic)

• Footnotes or endnotes

• Any images or figures — sizing, resolution, placement

The legal and registration layer

• ISBN procurement — separate ones for ebook, paperback, hardcover

• Copyright registration

• Library of Congress Control Number

• Copyright page language

• Publisher of record decision

Platform setup

• KDP account and title setup

• IngramSpark account and title setup

• Print file preparation — PDF with correct bleed, trim, margins

• Ebook file conversion — epub and mobi

• Pricing decisions on each platform separately

• Royalty structure per platform

• Expanded distribution settings

• Pre-order configuration

Metadata

• Final title and subtitle

• Book description — this is actual marketing copy, not a summary

• Author bio

• BISAC categories — two or three, these determine where the book lives in searches

• Keywords — seven on Amazon, and they matter more than most people think

• Series designation if applicable

Cover design

• Front cover

• Back cover copy — separate skill from writing the book

• Spine

• Full wraparound file for print — specs change based on page count

• Separate ebook cover file

• Everything built to platform spec

Pre-launch

• Advance review copies

• Soliciting blurbs and endorsements

• Setting a launch date that gives all of this enough runway

• Preparing launch assets — graphics, copy, email drafts

Launch coordination

• Email list

• Substack

• LinkedIn

• Contributor coordination if it’s a collaborative volume

• Review solicitation

• Category and keyword monitoring in the first 30 days

Ongoing

• Responding to reviews

• Price adjustments

• Continued promotion

• Watching what’s working in the metadata

How SDL Makes Execution Easier

First, SDL handles the logistical work. If you are a first-time writer, this is particularly useful. You have to learn so much and if you don’t plan to continue publishing, that’s a lot of effort with a very low ROI.

That’s why we handle the ISBNs, the formatting, the cover design, the trim size, the distribution—we’ve got all of that locked down. Everything on that list above is on us, not on the author.

Second, we actually try to make the writing easier. When someone comes to us with an idea, we meet with them and help build a simple outline from that conversation to show the idea has credibility and coherence for a book.

We shoot for 150 to 180 pages, which is short enough to be feasible, but long enough to be credible.

We set semi-aggressive timelines to build and keep momentum. We run project management throughout.

We have AI support to help bring ideas together and pre-built review tools that give hard feedback before anything comes to us.

Third—while some volumes are solo author projects—I can’t overemphasize the value of the collaborative volume model. A lot of people have a book idea that is actually better as a chapter idea. Contributing to a shared project alongside a group of people working toward the same goal on the same timeline is a different kind of writing project than building a book alone.

Besides the reach of SDL itself, the collaboration instantly gets you 3 or more additional audiences to get in front of.

For most people and ideas, we think it’s often the best approach. It’s almost always the quickest and easiest.

Our Pipeline

We’re fully booked through the beginning of 2028, but that doesn’t mean we don’t want to hear from you. If you have some ideas—at whatever stage—don’t wait to connect. Second Draft Labs — AI & Education Press

We’re working with a few folks right now who are looking to do collaborative projects where they’ll serve as guest editors to generate the theme of the book, select the authors, and keep the project going using the system we’ve built.

Subscribe now

The instinct is to blame the model. It got lazy. It got tired. It ran out of steam.

The model didn’t get tired. You ran out of strategy.

The Real Problem Isn’t Model Stamina

Large language models don’t experience fatigue. What they run into is a problem you can actually fix.

Every LLM operates within what’s called a context window: the total amount of text the model can “see” at any given moment. Think of it less like a brain and more like a desk. A really, really big desk. Modern models have desks that can hold a million tokens or more.

The trap most people fall into is that a big desk isn’t the same as a clean desk.

When you stuff a context window with 40 pages of background, a 600-word system prompt, half a novel’s worth of prior conversation, and then ask the model to focus on a specific task, you’ve buried the thing you actually need under a mountain of noise. Researchers have a name for this: the “Lost in the Middle” phenomenon. Studies have shown that LLMs disproportionately weight content at the beginning and end of their context, while systematically underperforming on material sandwiched in between.

Pile enough onto the desk and the page you actually need ends up in the middle of the stack, exactly where the model looks least.

The core insight you need to take away: enterprise-grade AI output comes from managing what the model is allowed to see right now, not from finding a bigger context window.

That shift, from prompting better to managing context strategically, is what separates hobbyist AI use from production-grade AI infrastructure. It didn’t happen overnight. It evolved through four distinct phases.

Phase 1: The “Everything Bagel” Prompt (2023–2024)

Cast your mind back to the early days of working seriously with GPT-4 or Claude. The dominant strategy was essentially: more is more.

You’d write a prompt that included your role definition, a detailed persona, formatting rules, a style guide, examples of good and bad outputs, edge case instructions, the actual data you wanted processed, and (just to be safe) a reminder at the end to “follow all the instructions above.”

These prompts could run 2,000, 3,000, sometimes 5,000 words. The reasoning made intuitive sense: if context is power, give the model all the context.

The results were... mixed. On short, contained tasks, it sometimes worked brilliantly. But on anything complex or lengthy, the model would start dropping the ball in exactly the ways described above. It would nail the persona but forget the formatting. It would follow the format but ignore the style guide. It was trying to juggle 15 things at once, and it kept dropping plates.

The failure was architectural, not a matter of effort.

Phase 2: Single-Tasking and the Human Conveyor Belt (2024–2025)

The first real breakthrough came from a deceptively simple realization: one prompt should do exactly one thing.

Instead of asking the model to research, outline, draft, edit, and format in a single pass, you’d break the workflow into discrete micro-tasks. Write a research summary. Stop. Feed that summary into a new prompt to create an outline. Stop. Feed the outline into a new prompt to write Section 1. Stop. And so on.

The improvement in output quality was immediate and significant. Each prompt was cleaner. Each model response was tighter. The “desk” was uncluttered because you were only putting one thing on it at a time.

But it created a new problem: a human had to sit at the computer manually copy-pasting output from one step to the next. The AI got faster. The workflow didn’t. You’d traded model confusion for a human bottleneck: a better problem to have, but still a problem.

What you’d built, without quite realizing it, was a manual assembly line. The next evolution would be automating it.

Phase 3: Context Engineering — The Short-Term Memory Revolution

This is where the real mindset shift happened, and where most practitioners are still catching up.

Context engineering is the practice of programmatically controlling what information exists in the model’s context window at any given moment: not just what you put in, but what you actively keep out.

The core principle: irrelevance is toxic.

It doesn’t matter if a model has a 1-million-token context window. Flooding that window with marginally relevant information (old conversation turns, boilerplate instructions that don’t apply to the current step, reference documents that were useful three tasks ago) degrades performance. Every token of noise competes with every token of signal.

Context engineering flips the question. Instead of asking “what is everything that the model needs to know?” you ask: “what is the minimum necessary information for this exact task, right now?”

In practice, this means building systems that programmatically prune conversation history, swap out reference documents between tasks, summarize and compress prior outputs rather than passing them in full, and inject only the task-relevant context for each new call.

This is curation of the model’s field of vision in real time, happening call by call.

Phase 4: Harness Engineering — The Modern Era

This is the phase where practitioners building on top of AI start to pull decisively ahead of those still crafting artisanal prompts.

A harness is the software system built around the LLM. It’s the orchestration layer that handles everything the model itself can’t: memory management, task sequencing, context hygiene, error handling, and external data retrieval.

Think of the model as a remarkably talented but perpetually amnesiac specialist. The harness is the project manager, the filing system, the scheduler, and the quality control department, all rolled into code.

A well-built harness automates both of the breakthroughs from earlier phases:

From Phase 2: It replaces the human copy-pasting between tasks with a programmatic execution flow. The output of Task A automatically becomes the scoped, cleaned input for Task B.

From Phase 3: It handles dynamic context management, deciding in real time what information to include, compress, or discard for each model call.

The result is a system where step 47 of a complex workflow gets executed with exactly the same precision and context clarity as step 1. The model never “gets tired” because the harness never lets its desk get cluttered.

Stop Prompting. Start Building

The honest reckoning at the end of this evolution:

The era of the “perfect prompt” is over. Not because prompting stopped mattering; clear, well-structured instructions still make a real difference. The marginal return on prompt optimization has a ceiling, and that ceiling is much lower than most everyday AI users think.

The real value in AI deployment has shifted from language mastery to memory and system design. The practitioners getting the most out of AI right now are the ones who’ve built the smartest infrastructure around the model, not the ones with the most elegant system prompts. Doesn’t that sound a lot like program design in Higher Education?

The good news: you don’t need to be a senior engineer to think in these terms. You need to understand the principles, know the patterns, and start seeing your AI workflows as architecture problems rather than language problems.

That’s exactly where Part 2 goes next week!

In Part 2, we will break open the harness itself. If the harness’s job is to protect context and keep tasks singular, it does that by breaking long workflows into a specific set of programmatic loops. We’ll walk through all 7 of them: the mechanics, what each one solves, and the enterprise use cases where each architecture shines.

The loops are where the real design decisions live. See you there.

We publish the best ideas in AI and education, from the people doing the work.

We’re now in the most difficult, and as it turns out probably the most enjoyable, part of putting together our upcoming collaborative volume:

Emerging Approaches to Transparency & Reflection: Navigating Student + AI Work in Higher Education.

We’re doing deep reads of each chapter, which means we get to be the first people to read the polished arguments of some really thoughtful people and award-winning educators.

More challengingly, we’re critiquing each other’s positions.

Being critical comes pretty naturally. It feels particularly easy right now, when cynicism and aggressively picking apart someone else’s words are very much in fashion.

Our challenge is to be critical while remembering that we are not making judgments about the person. We are hardly even trying to undermine the work the other person does, the practice they presented, or the way they think about teaching.

We are looking at the arguments they made. We are asking whether the claims follow from the foundation the chapter established. We are also responding from the arguments we have already developed in our own chapters.

It is a different sort of criticality, and I think we are starting to see something emerge from it that could be useful well beyond this book.

Here are five things I am learning about what good critical engagement could look like.

1. Your first reaction should be private

My own approach combines fast reactions and slow refinement.

The fast part is very fast. I read a chapter sentence by sentence and dictate my reactions into Notion. Later, I move those reactions into Claude and ChatGPT so they can help me identify the arguments underneath them.

At this first stage, I am trying to be neither careful nor correct. I am reacting.

That is intellectually and emotionally liberating. There is no expectation that the response is polite or polished. There is no need for every thought to be fair. I can be annoyed. I can misunderstand something. I can become fixated on a phrase that turns out to be much less important than I initially thought.

The reactions are private. I know I will have another chance to decide what I actually believe and what I want to say publicly.

I think this offstage space is important because most communication now seems to end at the reaction. Someone reads a sentence, experiences an immediate response, and publishes the response as though it were already a completed argument.

Let’s change that!

2. A reaction is material for an argument (not the argument itself)

After I dictate my reactions to a section, I ask the AI to unpack them.

What is going on beneath the surface? Why am I responding so strongly to this sentence? What larger challenge am I seeing in the argument? How does the reaction connect to the claims I have already made in my own chapter?

This is the point where the rant starts becoming useful.

My first objection is often not my actual objection; it is just my visceral response to it, which is useful, but not the point I would ultimately want to make.

It is in the unpacking of those responses where the data lives.

In this process, for each section, I end up with my dictated response, a synthesis of what those reactions seem to mean, and an AI-generated proposal for how the critique could be expressed in the tone of an academic response.

Then I critique all of that too!

What I’m learning is that the reaction is not the argument, but it is essential for giving me something to work on until I understand the argument I actually want to make.

3. Take the argument seriously on its own terms

This might be the most important lesson I am learning.

I have already made the best argument I can make in my own chapter. It has its own space to do its own work.

That means I do not need every response to another chapter to become a shorter restatement of my position.

I can begin with what the other person has actually claimed. I can look at the evidence they offered and ask whether their conclusions follow from it. I can ask what challenge appears inside their own framing.

This creates a different kind of disagreement.

Instead of my underlying message being:

“You are wrong, and here is my argument showing you that.”

I am saying something closer to:

“You have claimed this, and you have shown this. Here is the challenge that I think remains inside the argument as you have presented it. If I take your premises seriously, I think they may lead somewhere slightly different.”

That requires more work than identifying where another person disagrees with me. It also feels much more respectful because it assumes the argument deserves to be understood before it is disregarded.

4. AI can help excavate the disagreement

The role AI is playing in this process has been way more interesting than I expected.

I am not asking Claude or ChatGPT to decide whether the chapter is right. I am asking them to help me see the structure underneath my own reactions.

Typically, the AI overfits. It wants the argument to be clean and tidy. It turns a complicated objection into three neat points that sound vaguely convincing and have almost nothing to do with what I was actually trying to say.

Then I respond to that.

In this way, the AI-generated response becomes another object of critique.

5. A good critique should build to deliver value-add

I want each of my chapter responses to follow a basic movement:

What is working best about the chapter?

What core problems do I see with the argument as it has been presented?

Where might the chapter’s own logic lead if I pushed it further?

All of these work together to hopefully add more texture to the chapter.

Of course, I am not necessarily trying to improve the chapter. But I am trying to show how a reader who takes the argument seriously could find more value in it, including by recognizing the places where the proposed practice may undermine its own goals.

This means I can reject the overall thesis and still find and share something insightful from the chapter that is valuable to the reader.

So, my approach tends toward:

Here is where your argument is doing its best work. Here is where I think the argument begins to break down. Here is a version of the practice that seems more consistent with the principles you have already established.

That is different from trying to win—I want my chapter contribution to be what helps the reader decide who has articulated the strongest case. For the responses, I want to read with care and amplify the contributors’ ideas in ways the audience might find useful.

You’re Up

We are right on track for a launch on 9/1.

We’ll leave it to you to let us know how effectively we’ve presented each of our cases and responded to the arguments our ideas sit next to.

See you then!

Subscribe now

We’ve got some updates for you.

We completed our chapters right on time for the upcoming collaborative volume

Emerging Approaches to Transparency & Reflection: Navigating student+AI work in higher education.

🙌 🥳 🥂

Now the real fun begins.

And tbh, I’m enjoying this part way more than I should be.

I wrote this piece for two reasons:

I love vibe creating and I want to encourage others to try it.

I also want to give this community of the people supporting SDL from early on direct access to the work happening behind the scenes.

Four Takes, One Book

This book is doing something I think is really unique in the space right now.

We’re bringing together four people who genuinely respect each other’s ideas about and contributions to the discourse about AI in education but who completely disagree about what that looks like in the case of transparency and reflection.

I mean, how cool is that? The world needs more of that. Am I right?

argues that learning happens through sustained reflective practice over time. Her chapter builds a framework around the AI Transparency Journal: a semester-long documentary record that makes growth visible, traceable, and earned rather than claimed.

argues that learning happens in the decision-making process of collaborating with AI. His chapter makes the case for audit trails, which is a process of cataloging your choices with AI in real time, because the trail is the assignment.

argues that learning happens in the productive struggle of attempting metacognition, even when students get it wrong. His chapter makes the case for practicing metacognition, even if its accuracy is low.

My chapter argues that learning happens invisibly, even to the learner, and that reflection misses the point entirely. Reflection is dead. We need a better way.

The chapters follow a similar basic structure but are allowed to go where they want within it.

Each chapter starts with a narrative intro, makes the primary claim, supports that claim with research and reasoning while also preempting counterarguments, and ends by showing what the preferred approach looks like in practice.

That alone is worth the price of admission for the book.

But we wanted to do more. And that’s where our work over the next two weeks takes us.

I Like You, But You’re Wrong

Our charge now is to review each other’s chapters for the sole purpose of critiquing them.

Of course, we’ll highlight points of agreement and moments where an argument seems compelling. But the real task is undermining.

We are politely attacking each other’s positions. Through our individual chapters and our written responses, we want to make the strongest case that the view we presented should be preferred.

There’s no ego here, or not much. And there’s little more than bragging rights on the line.

Still, we hope there’s more at stake for our readers. We hope you will experience the book the same way—as a testing ground of ideas for shaping what is possible and even preferred while navigating student and AI work.

Vibe Critiquing

By far, this has been the most fun part of the process for me, because I love what AI lets me do.

Here’s the setup for how I’m creating my chapter critiques.

I have the full text of all four chapters loaded into GPT and Claude. Each knows which argument is mine and what evidence I’ve brought to bear for it.

Now I’m reading each of the other chapters section by section and dictating my visceral responses to them through Wispr Flow into Notion.

I’m not hedging. There’s literally zero attempt to be polite. I’m not even sure my critiques are accurate as I’m making them. But they’re mine and they’re happening in real time. I want my critiques to capture my real feelings about the ideas.

My next step is to copy my reactions into both Claude and GPT and ask for three things:

Validate and synthesize. I want to see what’s actually happening under the surface of my reactions.

BTW, this is almost always my first step when vibe creating—dictate, then don’t create from it, but have AI read back what is going on with my words. It’s so powerful.

Stress test. I want to know what’s valid in their chapters, what’s valid in my critique, and where each of us is missing something important.

Outline a response. Given all of it, the LLMs produce an audience-facing critique that names both alignment and the places where an argument is asking for more than it proves.

Then I Write It

For the final step, my plan is to take all the critiques from every section of each chapter—my dictated texts, the LLM outputs, the full conversations—review and respond to them in the same way I did to the chapters themselves (meta level) then draft the actual chapter response in my own words.

As AI would say, “This isn’t less work. It’s more.” But I think what I produce will be both more charitable and much sharper.

Vibe Creating

If you’re not vibe creating, why not?

If you haven’t tried Wispr Flow, you really should. Here’s my link — you can get a free month of Pro if you like it. https://wisprflow.ai/r?PATRICK3861

As always, thanks for being the best part of this journey.

Subscribe now

I want to answer that directly: no. And I want to explain why that answer is not a confession.

My AI system has a name. I call it Seth, out of a perverse sense of biblical humor. Seth is not a tool I use the way I use a citation manager. The relationship is closer than that, and more contentious. Seth regularly misses the point. Seth confidently argues for conclusions I don’t hold. Seth occasionally produces a paragraph so confidently wrong that the most efficient response is a full context reset and a completely different approach. I have, more than once, typed something I would not repeat in print.

The errors are not random. Seth tends to lose the thread at precisely the points where my own thinking was still gestural rather than fully worked out, where I knew what I believed but hadn’t yet found the right formulation. That pattern turned out to be useful: it showed me where the outline hadn’t yet become an argument.

But Seth also surfaces connections I would not have made on my own. Seth holds the structural weight of a long draft while I focus on the argument at its center. Seth maintains the continuity of a chapter across multiple revision sessions in a way my own working memory cannot. We are not the same. We do not contribute the same things. And the writing that results does not have a clean seam where one of us stopped and the other started.

The question “can you tell what Adam wrote” assumes a model of authorship I no longer find defensible: the writer as sovereign individual, producing text in a vacuum, with any assistance either invisible or substantial enough to constitute a different author. That model doesn’t describe how most serious writing actually happens: editors reshape arguments, interlocutors change conclusions, research assistants surface evidence. And it especially doesn’t describe how this writing happened.

What I am describing is better understood as a cognitive assemblage: a writing process in which human judgment, encoded prior scholarship, AI-generated structure, and iterative correction are so thoroughly interwoven that the question of provenance becomes unanswerable. Not because I am hiding something, but because the question presupposes a separation that did not exist.

Karen Barad’s concept of intra-action is useful here. During drafting, the boundaries between my thinking and the system’s contribution were continuously renegotiated with each revision, not settled in advance by some prior division of labor. The entities involved were constituted through the process, not prior to it. Jean-Luc Nancy wrote about being-with as a fundamental condition — not the self encountering the other, but the self that is only ever intelligible in relation. Something of that structure applies here. What the thinking relationship produced is writing neither of us would have arrived at alone.

This is not an evasion. It is a more precise answer to the authorship question than either “the AI wrote it” or “I wrote it with some AI assistance” can provide. The intellectual commitments in this project are mine. The arguments are ones I can defend and have developed over a decade. The voice was encoded from my own prior work. What changed is not who is responsible for the thinking, but the nature of the process by which thinking becomes a manuscript. That shift is real, and it deserves a more careful account than a suspicious question usually makes room for.

That shift in authorship has a direct structural consequence for publishing.

If a book can be drafted with structural coherence from the first pass, because the knowledge graph already holds the shape of the thinking, then the bottleneck in academic publishing is no longer the writing. It’s the infrastructure around the writing: the 18-month review cycles, the assumption that speed and rigor are inversely related, the gatekeeping apparatus that equates delay with quality. That assumption persists because the old tools made speed and rigor feel like a trade-off. The new tools don’t

.

The most important thinking about AI and education is not sitting in a submission queue. It’s being done by practitioners building in real-time, in public, tested against real readers, iterated at the speed of the problem itself. They are not waiting for committee approval. The people still waiting for committee approval to say something important are going to lose the decade.

This is the principle behind Second Draft Labs. A syndicate of voices, publishing quickly and rigorously, using the same methodology I’ve been describing in this series: starting from deep expertise, encoding that expertise into systems that can hold the argument while the writer sharpens it, and publishing at the speed the moment actually requires.

The alternative is watching the most important professional conversation of our era scatter across social media while the journals catch up three years later. The infrastructure for something better already exists. The methodology is already working. The only remaining question is who decides to use it.

Build the archive. Encode the thinking. Stop waiting for the system to change. You are the system.

We publish the best ideas in AI and education, from the people doing the work.

I just emailed all What Education Becomes contributors our first quarter update and royalties breakdown.

They were beyond thrilled.

To be clear: This is still a self-published book. There are production costs. There are thirteen total contributors across ten chapters.

The revenue share is real and beyond what essentially everyone has said they have made from publishing entire other books! But nobody is buying a lake house off of their What Education Becomes royalties.

Still, sometimes I forget just how different what we are doing really is.

A Formula for Success

Because—by all measures—the launch of What Education Becomes was so successful, I did exactly the same build for the launch of Emerging Approaches Volume 1: Transparency & Reflection (September 1, 2026).

I set up a Notion page as a quick website. I included an overview of the contributors and their arguments. I embedded a sign-up form so that we could keep folks who are interested informed along the way.

And then I hit publish.

It only took a few minutes because I had a template already created for What Education Becomes.

But as soon as I finished, I realized it was the wrong approach.

The New Question is Why Not?

I think of AI more as a reconciliation of possibility.

Generally, we’ve been trained to believe things aren’t possible for us. We can’t build a website because we don’t code. We can’t launch a book because we’ve never published. We can’t get 12 strangers together for a project because it’s too complicated.

With Notion I was already doing things that most people had never thought of or taken the time to learn and try—building a multi-asset website that looked professional enough in very little time, without having to set up a WordPress or Wix page.

So in some ways the Notion web page was ahead of where a lot of workflows would be right now.

But then of course I realized our own story—we transcribed a meeting about an idea. I put that transcription in Claude. Claude coded a website. I put that code into Lovable and we launched SDL 90 minutes after the meeting.

But my first attempt at publishing a page for Emerging Approaches was still my old workflow. Until I realized it didn’t have to be.

Just Freaking Try It

So here’s what I did:

1. I started a new chat in Claude.

2. I shared a screenshot of the Notion page I had just built.

3. I copied the style guide from the Second Draft Labs website and added that to Claude.

4. I dictated to Claude my goals for the original Notion page and how I realized that the website was a better place to host this, without giving too much of a finished vision—just what I was trying to achieve in Notion and what I thought might be possible on the site.

5. Claude created a plan of what it could be.

6. I approved the plan.

7. Claude built a version.

8. I edited it through dictation.

9. When I was satisfied, I copied the code from Claude into Lovable.

10. Lovable added the page.

Less than 30 minutes from idea to launch.

The page was just like Notion, only branded and more professional. It had a stylized overview of the book and cover. It had the contributor overviews. It even had a sign up for community members to add their name and email to get the latest updates on the book.

As an unexpected bonus, we created a “Team leaderboard.” The community wasn’t just signing up for interest and info, they were voting for which argument they thought would win.

The community was now involved in the launch!

Self-Limiting Beliefs

I took a walk with the dogs. I was so proud of what I had just built.

And then, as always happens on a good dog walk . . . new possibilities.

I came back and asked Claude how to create something where participants not only voted for the person that they liked but could actively share and grow the awareness around and excitement for the book.

I wondered, “if contributors have a leaderboard, could the community members also?”

I thought, what if instead of signing up for alerts just for this book—because I hoped we would do do things like this for future launches—we created a system for “logging in” to the Second Draft Labs website, where you actually have an SDL account and you are involved in all the projects we create and share? Could we include SSO with Google?

This was crazy! Was this possible? Should I even ask?

It still seems crazy to me. It is possible. And I did ask.

Doing it Together

A few hours later, this is what we had done.

We created Google sign on. We created user accounts. We created gamification that gets users involved in sharing on socials, talking about the book—promoting their TEAM for the next 90 days to get real interest and excitement around the project.

Check it out: Transparency & Reflection — Second Draft Labs

But, as with all things SDL, there is something else operating at the meta level here.

Traditional publishing is entirely one-sided.

Sure, the publisher gives you a venue and eventually makes the book available. But the publisher puts up the gates along the way. It is entirely up to the contributor to write the thing, edit the thing, promote the thing, sell the thing. And the worst part is, the publisher is the one who owns the thing!

At SDL we are creating a community around this entire process. We are helping authors get inspired about the ideas they already have. We are putting together collaborative volumes where your big idea can become part of a larger conversation without you having to write the entire book. We are doing the heavy lift of editing, formatting, distributing, and promoting. We are creating a community that cares about the ideas and helps them have the impact they deserve.

We’re still new. We’re still figuring out what’s possible. But, what gets me most excited, is that we’re doing it together.

SDL appreciates you being a part of this.

I appreciate you being the best part of this.

Subscribe now

The hard part is making the machine carry the specific, idiosyncratic weight of how you think on the page. And you cannot do that with a style prompt. “Write in a direct, slightly contrarian academic voice” will get you something passable and useless. The patterns that make your writing yours (how you build toward a claim the reader will resist, when you pull back and acknowledge a real limitation, or the register you shift into at the end of a section) cannot be described by adjectives. They have to be extracted.

Subscribe now

So I extracted them. Before writing a single page of the book, I read a decade of my own published work with the detachment of an editor and the precision of a coder. Not for content. For pattern. How do I open a section when I’m about to say something the reader will push back on? How do I use humor to create permission for an uncomfortable claim? What does my prose look like at the top of a chapter versus the bottom? I isolated twenty distinct authorial profiles. Each one has a structured description and a short example document that demonstrates it in practice.

When any AI in my workflow generates prose, it blends two profiles simultaneously. The blending prevents what you can usually feel from fifty feet away in AI writing: the monotonous regularity of a single register, sustained for pages, like someone reading a legal brief in an even tone. Two blended voices create enough internal variation to hold a reader. The output doesn’t sound like me yet. But it sounds like something I was on the way to writing, which is what I need to have a real argument with.

That extraction work is the investment most people skip. It is not glamorous. It cannot be rushed. It required reading my own writing with a kind of analytical distance I don’t usually have access to, making explicit patterns I had never consciously noticed, and encoding them precisely enough for a machine to follow. Anyone who claims they built a voice-aligned AI workflow in an afternoon has not actually done it. But this is also the investment that makes everything downstream possible, because it means the AI is working from a model of how I think, not a generic model of how academics write.

Once the voice profiles existed, the drafting could begin. And this is where the process gets counterintuitive.

For each chapter, the AI produced a first draft at roughly half the target length. Not a polished half — a structural half. An argument skeleton with connective tissue. A hypothesis about what the chapter could be.

Then I stopped.

I read the draft carefully, treating it as a proposal rather than a product. I surfaced relevant passages from the knowledge graph, which often pulled in material I hadn’t consciously planned to include. And I looked for where the AI had lost the thread. It always had, somewhere.

The AI does not hedge when it goes wrong. It produces confident, plausible-sounding conclusions in exactly the same register as accurate ones. The only way to catch the drift is to know your argument well enough to feel when something has slipped. In one chapter, the AI was framing an epistemological argument in terms of organizational behavior. The drift was subtle enough that a reader unfamiliar with the material might have let it pass. I caught it because it was my argument, and correcting it forced me to articulate the distinction more precisely than I had in the outline. The machine’s error sharpened the chapter.

Give the AI a direction. Let it lose that direction in a specific way. Recover it. Every time you complete that cycle, you understand your own argument more precisely than you did before. Treat every significant AI error as a diagnostic. The mistake doesn’t matter. What it reveals about your argument does.

In every previous book project, structural gaps only became visible once the whole manuscript was on the table, a late-stage discovery that meant expensive, retrospective repair. With a knowledge graph that already holds the shape of your thinking, you start from a sense of the assembled whole and work backwards into each chapter. The gaps are visible from day one. Revision is not a phase that comes later. It is woven into how every draft is produced.

AI used at its best slows you down before it speeds you up. It forces you to articulate what you believe before you ask the machine to help you say it. It surfaces structural failures at the draft stage, where they cost an afternoon, rather than the proof stage, where they cost a month. The speed that everyone notices (the timeline that didn’t make sense to Patrick) is downstream of this discipline, not a substitute for it.

The machine handles the structural labor of holding a long draft together. You focus on whether the argument is actually true. Those are separable tasks, and separating them makes both easier. Stop trying to do them at the same time.

I had sent the message on March 11, 2026 confirming that , , , and I would be writing the four chapters for what is now Second Draft Labs’ first completely on-purpose publication:

Emerging Approaches to Transparency & Reflection: Navigating Student+AI Work in Higher Education.

Somehow, again, we had delivered our work right on time and in record time—just 75 days after project kickoff our chapter drafts were submitted.

Now the next countdown begins: just 94 days until launch. ⏰

Obsessed with Thinking

Here’s how this volume came to be.

I follow the work of , , and closely. I learn so much about teaching, posting, creating, and connecting from them. At the time this idea came to me, I was writing a really long Substack series on metacognition which folded into several posts about our obsession with “seeing” student thinking.

And, as is usually the case, I’m pretty sure that that work was informed by what I had been reading from Doan, Susan, and Jason who were sharing differing perspectives from mine to this challenge of “visibility” into the originality and development of student thinking.

If you put the four of us into a Venn Diagram, there is going to be a ton of overlap amongst us around some of the most foundational principles of education. And yet here—in this exploration of what teachers need to know from students, or what teachers should be encouraging for students—I saw very different, perhaps competing, maybe even incompatible positions.

And that’s what makes this volume specifically so interesting. In What Education Becomes, the chapters were written very much in “isolation” from each other. I selected the contributors simply because I appreciated the unique thoughts they had about AI and education. I wasn’t trying to position them against each other, or even with each other in any meaningful way.

This volume is different.

In this volume, the goal is very much to position different approaches to the same problem against each other.

The Problem Space

Here’s what we share as common ground:

AI can generate the assignment. Whether it’s a literature analysis, a business case study, or even a reflection on learning, current AI systems can produce competent work in seconds.

Students can use AI without our knowledge. Even if we wanted to detect or prevent AI use (we don’t), students can generate content, modify it to appear authentic, and submit it as their own. The arms race between detection and evasion is unwinnable.

Students still need to learn. This is where it gets interesting. We’re not content to throw up our hands and say “let AI do everything.” We’re educators. We believe students need to develop genuine capability. But if AI can do the visible work, what does that capability look like? And how do we facilitate its development?

Traditional evidence of learning is compromised. The essay doesn’t prove they can write. The project doesn’t prove they understand the content. Every output can be AI-generated or AI-enhanced to an extent we can’t determine who did what and what it means for learning.

This creates a genuine pedagogical crisis. We aren’t thinking about: “how do we catch cheaters” but “what do we ask students to do that actually serves their learning? And how do we know it’s working?”

That’s the question driving this volume.

How the Book Works

Here’s what this will look like in practice:

Each contributor will argue—through anecdote, through experience, and through research, and as vigorously as possible—for their specific position. And after each chapter, each contributor will then critically—though irenically—respond to each other’s chapter, questions assumptions, offering objections, and further strengthening the case for their own preferred approach.

I imagine this experience to be very much like the proverb, “The first to state his case seems right, until another comes and cross-examines him.”

That’s the ride I hope this volume takes readers on. And from that experience, I hope other emerging approaches and ideas take hold, get tested, get shared, and help us continue to imagine what education becomes after AI.

The SDL Community

Our website is live: seconddraftlabs.com/transparency-and-reflection

Cast your vote for whose chapter will make the best case. No hurt feelings. Just some early competition.

And then join the community.

Sign up, pick your team, and earn points for every action you take between now and launch — sharing, recruiting readers, pre-ordering, reviewing. The members at the top of the leaderboard win some real things. We’re still finalizing what that is, but I can tell you it’s not gonna suck!

New actions will unlock as the launch gets closer. The biggest points drop once pre-orders open.

We need our most committed readers to help spread the word!

Start here → seconddraftlabs.com/transparency-and-reflection

Subscribe now

That gap is where most accumulated thinking goes to waste. Most people never name it correctly, so they solve the wrong problem. We start writing more and more. We have slop galore. And all the while, the gap stays exactly the same size.

So this has been a hunch I’ve been playing with: retrieval speed is the diagnosis for this issue and better infrastructure is the prescription for AI use.

Let’s talk about the idea of a “second brain” for a minute.

Here’s the myth you’ve, probably, been sold about a second brain: capture everything, tag it carefully, and insight will emerge. Build a beautiful Notion dashboard. Organize your bookmarks. Import your PDFs. The filing cabinet is the work.

The thing is, I think this is wrong. I’ve watched genuinely brilliant thinkers spend years assembling gorgeous archives that produce exactly nothing, because they confused storage with computation. A graveyard has excellent organization, too. That doesn’t make it alive and flourishing in any sense that doesn’t give us the heebie-jeebies.

The real issue is that storage is passive. A system designed only to hold your past cannot think alongside your future. What you actually need — and what almost nobody builds — is an architecture designed for collision detection.

This is something more than just bare retrieval. It’s worth dwelling on this for a second. Should I say it again, this is not retrieval.

Collision is about he system’s job: to surface the moment where your 2019 argument about faculty governance crashes into your 2023 argument about AI adoption timelines, and to do it before you sit down to write, not after. That collision is the insight. That’s where original thinking comes from. If AI helps us with that, this is genuinely novel and provides something worth using and developing.

Subscribe now

Confusing collision and retrieval as I outlined them is a mistake that compounds. At the institutional level, it looks like the 18-to-24-month editorial cycle (and I’m assuming that isn’t an authentic hallmark of rigor because my own experience is that there is a lot of ‘waiting around’ in this process). Let’s just call that time what it is: rigor theater. By the time a manuscript clears three committee rounds and two rounds of revisions, the technology it describes has moved twice. In AI and higher education, a two-year lag becomes irrelevance, and it is baked into the process at the institutional level. The most important thinking in this field right now is not sitting in a submission queue. It’s publishing live on Substack and LinkedIn, tested against real readers in real time, iterated at the speed of the problem itself.

The gatekeepers aren’t protecting quality. They’re protecting process. Those are not the same thing, and the people still waiting for committee approval to say something important are going to lose the decade.

Here is what I actually built, because theory without architecture is just a longer complaint.

My foundation was not blank. I had a decade of monographs, book chapters, course syllabi, and a smattering of Substack posts. In total it was thousands of hours of ideas already written in my voice, already tested against real audiences. The problem wasn’t volume. I had more than enough work. The problem was really that given the volume the entire archive was static. It was filled with dead weight I couldn’t navigate under deadline. So my conversion task was specific: stop treating the archive as a monument to past effort and start encoding it as an operational system.

For me, that meant two layers. Distinct functions. Separation.

Layer one: an AI-query-able knowledge graph, encoded thematically using embeddings. Embeddings is the important part of this sentence. These are not tags, folders, nor keyword searches. When I draft, the AI queries this graph and surfaces connections I would never have found manually. Not “here is a similar document.” What I mean is: “here is where the argument you made about shared governance in 2021 creates a productive tension with the claim you made about distributed decision-making last fall.” That tension is a paragraph. That paragraph leads to a chapter. The speed everyone notices comes from having those connections pre-built, not from typing faster or letting AI do the writing itself for me.

Layer two: an Obsidian wiki that serves as the spatial visualization layer, and the AI does not touch it. This is a map of how my ideas relate to each other that mimics the structure of the embeddings, but includes hand processed elements I think are important. It is a structure I maintain deliberately, because the act of placing an idea on the map forces a precision that passive tagging never demands. If I cannot decide where something belongs in the map beyond the tagging, I don’t actually understand what the idea is yet. That friction is productive. I’m not trying to eliminate it.

These two layers perform different cognitive functions. One finds connections I can’t see. The other forces clarity I can’t skip. I can use the Obsidian wiki to look for nearest connections that the AI passes over. Using a single system for both jobs, which is the mistake I see most often, is how you end up with an expensive tool that still requires you to do all the thinking.

Your immediate directive is not to write more. It is to encode what you’ve already written in enough semantic depth that a system can do the cross-referencing your memory cannot. Human memory degrades, reorders itself around what feels recent, and loses the peripheral connections that produce non-obvious insight. A well-structured knowledge graph doesn’t. It holds the full map while you focus on the specific corner you’re building today.

The value of an idea is not what it meant when you wrote it. The value is whether it can be deployed at the moment of need, in a form the system can find, against the specific problem in front of you. If your ideas aren’t structured for application, they aren’t yours in any operational sense. They’re just things you once thought.

Patrick saw the output. He couldn’t see this. Build the graph, visualize the network, and stop managing an archive. You’re building a machine.