Turn a single photo into a short Gemini Omni Flash video that feels as though a real person filmed it.

Copy this prompt

Turned out well, the right side looks different than in real life, but it holds for a good 7 seconds.

Pretty cool to think how you can turn static photos into videos, with some simple steering.

Phil

I used one photo of Acapulco Bay to make a 2.5D parallax shot.

What is a 2.5D parallax shot?

A 2.5D parallax shot makes one flat photo feel as if the camera is moving through a real space.

It is not full 3D. The photo is treated as a few depth layers: things close to the camera, things in the middle, and things far away. When the camera moves sideways, the close layer appears to move fastest, the middle layer moves a little less, and the horizon hardly moves at all.

Your brain reads that difference as depth. It is the same visual clue you notice when looking out of a moving car: a fence rushes past, nearby trees slide across your view, and distant mountains barely seem to move.

Image to explain it

The editing skill, in plain English

The skill is called parallax.

You are faking camera movement by making different parts of one image move at different speeds. Things that should be nearer to the viewer move more. Things that should be far away move less.

For this image:

• The palms and bougainvillea are closest, so they slide fastest.

• The bay is further away, so it slides more slowly.

• The islands and horizon are furthest away, so they barely shift.

What I used

One landscape photo from my place in Acapulco: plants and bougainvillea in front, the bay in the middle, then islands and boats further out.

The prompt

Phil

WIRED reported this week that Anthropic added a safeguard before the US government lifted restrictions on Claude Fable 5 and Mythos 5. The guardrail can reroute some restricted requests to the older Opus 4.8 model. Access blocked, guardrail adjusted, model back online. Messy, but there it is.

Prompt writers can leave the politics alone for a minute. The useful part I want to talk about is the boundary problem.

Claude Fable 5 can be impressive and irritating because it does not always wait in the way a person expects. Anthropic’s Claude Code docs describe an agentic tool that reads a codebase, edits files, runs commands, and works across development tools. The best-practices page puts it even more plainly: Claude Code can read files, run commands, make changes, and work through problems while the user watches, redirects, or steps away.

When you start desinging prompts, you want to consider the frame.

A problem statement is not always an instruction.

The user may be thinking out loud. They may want a diagnosis. They may be checking whether their read of the situation makes sense before they choose the next move.

The agent should not treat every sentence as a trigger.

First boundary:

When the user is describing a problem, asking a question, or thinking out loud, the output is an assessment.

Not a patch.
Not a draft.
Not a branch.
Not some tiny “while I was out” improvement.

An assessment.

The agent reports what it found. It names the likely cause. It says what evidence is weak or missing. Then it stops.

Commands need a harder line. Restarting a service, deleting files, editing config, resetting a cache, changing environment variables, renaming branches, or modifying generated assets should not happen just because the agent recognises the shape of a problem.

Pattern matching is not evidence.

A failed build might be a dependency issue. It might be a bad import, a stale lockfile, a version mismatch, a missing environment variable, or a broken test pretending to be a build problem.

A slow app might be cache. It might be a database query, a network call, a memory leak, or some third-party service dragging its feet.

The agent should be able to pause and say:

“I recognise the pattern, but I don’t have enough evidence to act yet.”

Teachers know this problem too well. A student gives a half-right answer, and you can usually guess what went wrong. Maybe they copied the method. Maybe they misunderstood the word in the question. Maybe they guessed and got lucky.

The pattern helps, but it is not proof, just like a test isn’t proof of the human’s knowledge.

Use this if you are struggling to work out why a student is not getting a concept:

“Assessment mode first. The exact learning goal is <Learning Goal>. This is the student’s answer <Student Answer>, but do not improve it yet. First tell me what the student seems to understand, what they may be guessing, what misconception might be blocking them, what evidence supports that, and what evidence is missing. Then suggest what I should check next. Stop there unless I ask you to rewrite, reteach, or create a task.”

Example:

Learning Goal: Students will be able to explain how the greenhouse effect works by describing the relationship between solar radiation, the Earth’s surface, and atmospheric gases.

Student’s Answer: “The sun hits the Earth and warms it up, and then the heat stays trapped inside the atmosphere because of pollution like smoke and gas. The ozone layer is getting holes in it, which lets too much sun in, and that’s why the planet is getting hotter.”

“Assessment mode first. The exact learning goal is <Students will be able to explain how the greenhouse effect works by describing the relationship between solar radiation, the Earth’s surface, and atmospheric gases.>. This is the student’s answer <“The sun hits the Earth and warms it up, and then the heat stays trapped inside the atmosphere because of pollution like smoke and gas. The ozone layer is getting holes in it, which lets too much sun in, and that’s why the planet is getting hotter.”>, but do not improve it yet. First tell me what the student seems to understand, what they may be guessing, what misconception might be blocking them, what evidence supports that, and what evidence is missing. Then suggest what I should check next. Stop there unless I ask you to rewrite, reteach, or create a task.”

Then the teacher can choose the next move:

“Give me a reteach explanation.”

“Write three diagnostic questions.”

“Show me the misconception.”

“Make a simpler example.”

“Rewrite the task.”

“Leave the answer as it is and just give feedback.”

Phil

Think of Claude as Buzz Lightyear.

Advanced. Polished. Probably a bit too confident.

But even Buzz needs a briefing.

He can have all the tools in the world and still not know what Woody actually needs unless someone tells him.

That is the simplest way to understand a Claude Skill.

A Skill is a note you leave for Claude so it has useful context before the work starts.

Not coding.

Not some giant AI system.

Just a reusable briefing.

Something that says:

Here’s who I am.
Here’s what I’m working on.
Here’s how I like help.

That’s it, mostly.

The easiest analogy

A Skill is like the short briefing document you would give a new assistant on their first day.

You would not hire someone, sit them at a desk, and say:

Help me.

Well, you could.

But they would have to ask a pile of questions before doing anything useful.

They would need to know what you do, who you help, what kind of work matters, what tone sounds right, what annoys you, and what “good” looks like.

A Claude Skill does that job.

It gives the AI the background information you do not want to explain every single time.

Buzz without a Skill

Woody says:

Help me write an email.

Buzz replies:

Certainly. Who is the audience? What tone would you like? Should this be formal or casual? What is the goal? How long should it be?

Buzz is not being difficult.

He just does not know the room yet.

He does not know Woody’s usual style. He does not know the audience. He does not know whether Woody wants something quick, warm, blunt, funny, careful, or professional.

So the conversation starts with setup.

Again.

Buzz with a Skill

Now imagine Woody has left Buzz a note.

I run a beginner-friendly newsletter about AI. My readers are smart, but not technical. I like plain English, simple analogies, short paragraphs, and practical examples. Avoid jargon unless you explain it.

Now Woody can say:

Help me write my newsletter.

And Buzz has a much better starting point.

He already knows the audience.

He already knows the tone.

He already knows not to start throwing in technical words just because he can.

That is the point of a Skill.

It saves you from repeating the same context every time you open a new conversation.

What a Skill looks like

A Skill does not need to sound impressive.

In fact, the more impressive it tries to sound, the less useful it often becomes.

A simple Skill might look like this:

I am a small business owner who uses AI to write newsletters, organise ideas, and create simple marketing content.
I prefer clear explanations, practical examples, and short drafts.
Avoid technical jargon unless it is needed, and explain it in plain English when it appears.

That is enough to start.

It gives Claude a job shape.

It tells Claude what kind of person it is helping, what kind of work is common, and what kind of response will probably be useful.

Skill vs prompt

A prompt is the thing you ask right now.

A Skill is the background note that helps Claude understand how to answer.

For example:

Prompt:

Write a welcome email for my newsletter.

Skill:

My newsletter is for beginners learning AI. My tone is simple, practical, and friendly. I like short paragraphs, examples, and no unexplained jargon.

The prompt is the task.

The Skill is the briefing.

Or even simpler:

Prompt = what to do.

Skill = how to understand you while doing it.

The 3 things every useful Skill needs

You do not need a complicated system.

Most useful Skills contain three things.

1. Who you are

Tell Claude your role, project, or situation.

For example:

I write a beginner-friendly newsletter about AI for small business owners.

This gives Claude the world you are working in.

2. How you like things done

Tell Claude your preferences.

For example:

Use plain English, short paragraphs, simple examples, and a calm, practical tone.

This helps Claude avoid giving you the wrong kind of answer.

3. What help you usually need

Tell Claude the tasks you often ask for.

For example:

Help me write newsletter drafts, social posts, simple explanations, and reader exercises.

This helps Claude become useful faster.

Not perfect.

Just less blank.

A simple formula

Here is the beginner version:

I am ________.
I work on ________.
When helping me, please ________.

That is enough for a first Skill.

You can always improve it later.

Most people overthink this part. They try to write the perfect AI instruction and end up writing something stiff.

Start small.

Write the note you would actually give a new assistant in their first days.

Copy-paste template

Use this:

I am a ________ who works on ________.
My audience, customers, or readers are ________.
When helping me, use ________ and avoid ________.

Example:

I am a newsletter writer who helps beginners understand AI.
My readers are curious, but not technical.
When helping me, use plain English, simple analogies, short paragraphs, and avoid jargon.

That is basically your first Skill.

Not fancy.

Useful.

Reader action step

Write a 3-sentence brief about yourself as if you were onboarding a new assistant.

Use this:

I am ________.
I am working on ________.
When helping me, please ________.

You do not need to build anything complicated today.

Just leave Claude a better note.

Phil

Here’s a free trigger sentence you can paste into any Claude Skill right now:

Use this Skill when I ask about meal planning, dinner ideas, what to cook, grocery lists, or say "I don't know what to eat."

LEARNING GOALS

After reading this blog, you will:

1. Know what a trigger sentence is and where it goes

2. Be able to spot a bad trigger sentence in five seconds

3. Write your own trigger sentence for a Skill you actually use

THE EXPLANATION

A Claude Skill is a saved set of instructions. The trigger sentence is the part that tells Claude when to use them. Without a good trigger, your Skill is a T-Rex in a paddock with the fence off, it either never shows up, or it stomps into every conversation uninvited.

Dr. Malcolm had a line in Jurassic Park: “Life finds a way.” Claude finds a way too. If your trigger sentence is vague, Claude will guess. And Claude guessing is chaos theory in action. Small input, wild output.

Here’s the difference.

Vague trigger (bad): “Use this Skill when I need help with life stuff.”

What does “life stuff” mean? Taxes? Dating? Whether to buy oat milk? Claude has no idea. So it either activates constantly or never.

Specific trigger (good): “Use this Skill when I mention food, dinner, cooking, groceries, or what to eat.”

Now Claude has a checklist. Five clear words. If you say any of them, the Skill wakes up. If you don’t, it stays asleep.

How Claude reads your trigger

Claude scans your message. It checks the trigger sentence against what you said. If it sees a match, a word, a topic, a phrase, it loads the Skill. If not, it carries on like the Skill doesn’t exist.

That’s it. No magic. No AI mystery. It’s a switch with words on it.

Three trigger sentences you can steal

For a meal planning Skill: “Use this Skill when I mention food, dinner, cooking, groceries, meal prep, or say ‘what should I eat.’”

For a lesson planning Skill: “Use this Skill when I mention lesson plans, classroom activities, learning objectives, or a year level like Year 5 or Year 8.”

For an email-writing Skill: “Use this Skill when I ask to write an email, reply to a message, or say ‘draft this.’”

FREEBIE

Copy-paste this template into any Skill:

Use this Skill when I [action] or say [exact phrase].

Trigger words: [word 1], [word 2], [word 3], [word 4], [word 5]

Trigger phrases: "[phrase 1]", "[phrase 2]", "[phrase 3]"

Do not activate this Skill for general conversation, greetings, or unrelated topics.

Fill in the brackets. Save the Skill. Done.

READER ACTION STEP

• Pick one Skill you’ve already built (or want to build)

• Write the trigger sentence using the template above with at least 5 specific words and 2 specific phrases

• Test it: open a fresh chat, say one of your trigger words, and check if the Skill activates

SIMPLE SUCCESS CHECK

You got it if you can say one trigger word in a new chat and watch your Skill switch on without you asking for it.

🔧 Prompt Engineering Tip: When you write trigger sentences for your own Skills, run them through Claude first with this prompt:

Here's my trigger sentence: [paste it]. List 5 messages a user might send that SHOULD activate this Skill, and 5 that SHOULDN'T. Tell me if my trigger is too vague, too narrow, or just right.

This catches the chaos before it hatches.

Phil

Claude acts like it’s never met you before because, by default, it doesn’t carry your personal context from one chat to the next. The fix is to teach it your basics once, so you stop re-explaining the same things every week.

What this post does

This post should make the problem feel familiar first: the endless loop of re-briefing, re-prompting, and repeating yourself. It then turns that pain into a simple habit — teach Claude once, then let that context shape future answers.

Freebie

Use these three copy-paste prompts today:

Act as my specialized consultant. Here is my context: [paste context]. From now on, whenever I ask for help, synthesize your answer through the lens of this context.

Before you execute any task, tell me what you need to know about my constraints or preferences to get this right on the first try.

I keep re-explaining myself to AI. First, ask me the 5 most important things you should remember about how I work, what I care about, and what I need.

The AI broken down

Claude “forgets” you because each new conversation starts without the background you have in your head. It can remember some things over time, I think it’s always better to have a document of your contextual requirements around, so that if you use other LLMs, it can transfer easily. That means your role, tone, audience, preferences, and repeated constraints all have to be rebuilt from scratch unless you store them in a way Claude can reuse.

A better mental model is this: instead of explaining yourself forever, you’re creating a short briefing that Claude can use every time. It’s the difference between giving directions once and having to redraw the map for every trip.

A simple way to start is to list the three things you explain most often. For many people, that will be tone, audience, and format. Once you write those down, you’ve already built the core of your first Skill.

Reader action

Write down 3 things you re-explain to Claude every single week. That list becomes the raw material for tomorrow’s setup.

“Constraint setting” sounds like it belongs in a software developer’s handbook. It doesn’t. It means writing a list of what you don’t want, right inside your prompt, before Claude has a chance to guess wrong.

Most people only tell Claude what they want. They leave the rest to the AI. Claude has no idea what to leave out.

❌ What most people do

The prompt:

Analyse this professional development proposal for our district.

Claude returns four pages. There’s a section comparing alternative providers, you didn’t ask for that. There are cost projections running out eighteen months. The language reads like a policy brief: formal, hedged, the kind of writing that makes a principal’s eyes glaze over in thirty seconds. Nothing is wrong. Nothing fits either. You spend twenty minutes editing it back to something usable, which is the exact problem you were trying to avoid.

✅ What to do instead

Add a DO NOT section. Write it before the content, not after.

The prompt:

Analyse this professional development proposal for our district.

DO NOT:
- Use technical jargon without a plain-language explanation
- Include financial projections beyond six months
- Name specific competing providers or products
- Recommend anything requiring more than $10,000 to implement
- Write in a formal or academic tone — this goes to a working principal,
  not a policy board

Focus only on practical implementation implications.

Three things you can try right now

1. Rerun your last Claude output with a DO NOT list added. Pull up the last professional document you generated through Claude. Add three constraint lines and run it again. Notice what disappears from the response.

2. Build a constraint list for your grant summary workflow. If you write federal or foundation grants, tell Claude upfront: no unverified statistics, no budget figures outside the NOFO parameters, no language that doesn’t appear in the original grant requirements. You’ll get compliance-ready language the first time through, not after two rounds of cleanup.

3. Add a scope limiter to your curriculum review prompts. When analysing a scope and sequence, add: “Focus only on ELA strand alignment. Do not comment on pacing unless it directly relates to standards coverage.” Claude goes deep on one thing rather than shallow on everything.

How to save it for next time

Drop this into your Claude Project instructions. It runs quietly in the background every time you open a conversation:

When I ask you to analyse any document — a proposal, a curriculum plan, 
a grant draft, a scope and sequence — apply these constraints by default:

DO NOT:
- Use education jargon without a plain-language explanation
- Include financial projections beyond 12 months
- Name specific third-party products, vendors, or competitors
- Recommend anything requiring more than [your budget ceiling] to implement
- Write in an academic or policy tone — my audience is practitioners,
  not boards

Focus on practical implications only. If something falls outside that 
scope, flag it in one sentence and move on.

Replace [your budget ceiling] with the real number for your district clients. That one substitution means Claude is already working inside your actual constraints before the conversation starts.

Add a DO NOT section to your next Claude prompt today.

Built with AI Blueprints for Education × PBL Future Labs — Phil

Friday, 5:21pm Eastern. A letter arrives at Anthropic’s offices, and somewhere in the world, a few thousand workflows stop mid-sentence.

The argument started immediately. Was it Trump admin retaliation for Anthropic’s refusal to strip guardrails from their models for autonomous weapons use? A pre-IPO marketing stunt? A genuine national security call? Two thousand comments, most of them missing the point.

The point is that they could.

One letter. One Friday evening. No warning, no due process, no appeal. The model you’d been using to fix your codebase, run your research, build your clients’ workflows ..gone. Not because it stopped working. Because someone in government decided it should stop working.

I wrote about this. Last year.

Here’s the short version.

The architecture of dependence

Every tool you use that lives on someone else’s server, governed by someone else’s terms, subject to someone else’s jurisdiction, you’re borrowing it. Not owning. Not even renting, because renters have rights.

In The Great Bifurcation, I wrote about standing under my kitchen sink with a pile of abstract knowledge from videos and AI guides, useless the moment my hands were on the pipe. “I was an editor of information,” I wrote, “but I wasn’t a plumber.” Friday’s ban is the macro version of that. Thousands of people thought they had a tool. They had access to a tool. Different thing entirely.

This happened before. In the 1990s, the US government classified strong encryption as a munition. Netscape shipped weakened SSL internationally for years. Developers printed PGP source code on paper and mailed it overseas, because the law allowed exporting a book but not software. The government eventually backed down, but only after years of legal battles, only after the internet had decentralised enough to make enforcement impossible.

We haven’t decentralised AI. We did the opposite. We handed it to four companies, built critical workflows on their APIs, and called it progress.

In The Sculptor’s Error, I called this out directly: “we are trading the skill of thinking for the convenience of an answer.” Friday made the material version of that visible. When Fable went dark, you didn’t just lose a model. You lost everything you’d built on the assumption it would stay.

Every district that approved an AI tool without asking “what happens when we can’t access this?” just watched the answer arrive. Every consultant who built their practice through someone else’s API got the same lesson. The kill switch was never a hypothetical.

The kill switch was always there.

Friday was the first time they used it.

- Phil

Start with an Empty Square

Prompt idea:

Create a simple square image with a plain white background and nothing inside it.

Then keep adding one object at a time:

1. Find and add one red flower in the bottom-left corner. Use the prompt - extract the flower and add it to the bottom left of the square.

2. Ask it to add be inspired by the red flower and add 3 extra plants similar to that flower on each corner. Use the prompt - add 3 extra plants similar to that flower on each corner.

Goal: this shows how a blank image becomes a controlled composition.

Meta-prompting. Say it out loud and it sounds like it belongs in a computer science degree. It doesn’t. Essentially you ask Claude to look at your prompt and tell you what’s wrong with it before you run it.

You use AI to fix how you talk to AI.

❌ What most people do

Here’s a prompt most education consultants have written at least once:

Write a professional development session on AI for teachers.

Claude returns a generic 60-minute outline with five vague sections and an icebreaker.

✅ What to do instead

Before you run your main prompt, run this one first.

Prompt Optimisation Template:

I'm trying to get better results for [specific task].
Here's my current prompt: [paste your prompt]

Please:
1. Identify what might be unclear or missing
2. Suggest specific improvements
3. Rewrite an optimised version
4. Explain why your changes would work better

What comes back is a numbered diagnosis, four to six things Claude flagged as missing or underspecified.

Task Analysis Helper:

I want to prompt an AI to [describe your goal].
What information would the AI need to do this well?
What format should I request for the output?
What constraints or guidelines should I include?
Create a complete prompt template for me.

This returns a full, reusable prompt template. Best for recurring deliverables.

Three things to try right now

1. Find the last prompt you wrote that gave you something flat. Paste it into the Prompt Optimisation Template above. Read the diagnosis.

2. Pick a document you’ve built from scratch more than twice, a standards-aligned unit summary, a consultant proposal, a scope-and-sequence overview. Run it through the Task Analysis Helper and save the template it gives you back.

3. Run your original prompt. Then run the rewritten version Claude gave you. Put both outputs side by side. That gap is your benchmark. Every prompt you write after this gets measured against it.

Open Claude, paste a prompt that gave you something forgettable, and add [OPTIMISE].

Phil

“Retrieval-Augmented Prompting.” Say it out loud and it sounds like something a developer drops into a meeting right before everyone looks at their phones. Here’s what it actually means: you give Claude your specific document, your audit, your student data, your district report, and you tell it which parts to pull from your material and which parts to pull from its own general knowledge. Two sources. Kept separate.

❌ What most people do

Paste a document. Ask a question.

Here's our curriculum audit. What should we do next?

The problem is that Claude doesn’t know where to draw the line between your audit and everything it already knows. You don’t know either. You might get back recommendations that sound confident and relevant but have nothing to do with what’s actually in your document. If a curriculum coordinator asks where the advice came from, you can’t point to anything.

✅ What it looks like when you separate the sources

Prompt:

Based on the curriculum audit I've pasted below, please:

1. Summarise what type of document this is and its main purpose
2. Pull out the key findings — from the document only
3. Flag any gaps the document doesn't address — document only
4. Compare those findings to common practices in standards-aligned 
   curriculum design — use your own knowledge here
5. Suggest next steps — draw on both sources

Label each section clearly. Keep document content and your own 
analysis separate throughout.

[Paste your audit here]

Three things you can try right now

1. Paste last semester’s student achievement data into Claude. Add this line at the top of your prompt: “Use only this data for trends. Use your general knowledge for benchmarks and comparisons. Label each section clearly.”

2. Drop in a school improvement plan and ask Claude to pull the stated goals from the document only, then separately tell you which of those goals are most commonly prioritised in districts facing similar challenges.

3. Take any professional learning session you’ve run. Paste the participant feedback. Ask Claude to identify recurring themes from the feedback only, then compare those themes to what research suggests educators typically need most. I often turn it into slide decks. This is why metacognition is important.

How to save this in your Claude Project

If you regularly work with school data, audit reports, or client documents, add this to your Claude Project instructions. You type it once and it runs on every conversation in that project.

When I share a document or a data set, always structure your 
response in two parts:

Part 1 — From the document: findings, themes, and gaps drawn 
only from what I've shared.

Part 2 — From your knowledge: comparisons to best practice, 
benchmarks, or research.

Label these two parts clearly in every response. Always tell me 
which source you're drawing from before you switch.

Open Claude, find a document you’ve been putting off, and run the prompt above.

Phil

Few-shot prompting sounds like it belongs in a computer science paper. It doesn’t. It means showing the AI examples of what you want instead of describing it. You paste in two or three of your own emails, summaries, or reports, then ask the AI to write another one in the same style. The “few” is the number of examples. The “shot” is each one.

❌ Most people start here:

Prompt:

Write a professional email declining a meeting request. 
Make it polite but brief.

The AI has to guess what “polite,” “professional,” and “brief” mean to you specifically. Whether you use first names. Whether you say “thanks” or “I appreciate your understanding.” Whether your emails run two lines or six. It has no idea. So it picks the safe middle of everything, and what comes back sounds like a template, not you.

✅ This works:

Prompt:

Here are examples of how I write professional emails:

Example 1 — Rescheduling:
Hi Sarah, I need to move our Tuesday meeting due to a conflict. 
Would Thursday at the same time work? Thanks!

Example 2 — Declining a request:
Hi John, I won't be able to make the meeting tomorrow. 
Could we find an alternative time next week? I appreciate your understanding.

Now write an email declining the Friday budget review meeting. 
Match the style of these examples exactly.

Three things to try this week:

1. Match your workshop summaries. Paste two summaries you’ve already written for district clients. Then give Claude your rough notes from your last session and ask it to write a third in the same style. You get your own voice back, not a generic version of it.

2. Set a standard for your learning objective language. If you write curriculum documents, find two examples of how you phrase objectives. Paste them in. Ask Claude to write the next five in the same format. What takes an hour of consistency-checking takes thirty seconds.

3. Draft post-workshop follow-up emails. Find two emails you’ve sent after sessions that led to a follow-up booking. Paste them as examples. Then give Claude three bullet points about what happened in your last session and ask it to write the follow-up. The result reads like you wrote it.

How to save it for next time:

If you use Claude Projects, you can store your example emails inside the project instructions once and never paste them again. Drop this block straight into your project custom instruction setup:

When I ask you to write professional emails or summaries, 
match the style of these examples exactly:

[Paste Example 1 here]

[Paste Example 2 here]

Do not explain your approach. Write in this style by default.

Every new conversation in that project will start already knowing how you write. No repeating yourself. No rebuilding the setup from scratch.

Open Claude now, paste two emails you’ve actually sent, and ask it to write the one you need today.

Built with AI Blueprints for Education × PBL Future Labs - Phil

Welcome! If you’ve been trying to work out this wacky world of artificial intelligence recently, you’ve probably heard a lot of buzzwords flying around. Two terms that pop up constantly are Markdown files and AI Agents.

Because they are often talked about at the same time, it is incredibly common to wonder: Wait, are these the same thing?

The short answer is no, they are completely different! But the exciting part is how beautifully they work together.

To help make sense of it all, let’s look at a simple analogy: The Recipe vs. The Chef.

What is a Markdown File? (The Recipe)

A Markdown file (often ending in .md) is simply a text document.

Imagine you are writing a recipe on a plain index card. You want the title to be big, the ingredients to be a bulleted list, and a few important words to be bold. Markdown is just a super easy way to format text using regular keyboard symbols instead of clicking buttons in Microsoft Word.

For example, typing # Hello makes a big title, and typing * Apples makes a bullet point.

The Key Takeaway: A Markdown file is static. It is just information sitting on a page. It cannot think, it cannot move, and it cannot bake a cake on its own. It is simply the recipe.

What is an AI Agent? (The Chef)

An AI Agent is a smart, active computer program designed to get things done.

Think of an AI Agent like a highly skilled (hopefully adaptive) robotic chef in your kitchen. If you tell the chef, “I want a chocolate cake,” the chef will look in the pantry, gather the ingredients, mix them up, and bake the cake for you. AI Agents can read data, search the internet, make decisions, and create things based on your instructions.

The Key Takeaway: An AI Agent is dynamic. It acts, thinks, and solves problems. It is the chef.

The Core Differences at a Glance

Here is a quick cheat sheet to remember the difference:

• What it is: Markdown is a document. An AI Agent is a program.

• What it does: Markdown holds text. An AI Agent takes action.

• Brainpower: Markdown has zero brainpower (it’s just text!). An AI Agent has artificial intelligence to figure things out.

How They Work Together (A Step-by-Step Guide)

You might be wondering, “If they are so different, why do people talk about them together?”

Great question! AI Agents love Markdown files. Because Markdown is plain, clean text without any messy background code, it is the easiest way for an AI to read your instructions.

Here is how you can practically use them together:

Step 1: Write your instructions in a Markdown file

Open a simple text editor (like Notepad on Windows or TextEdit on Mac) and write a set of instructions for your AI. Using Markdown helps organize your thoughts for the AI.

It might look like this:

# Social Media Task
Please read the latest news about space travel.

## What I need you to do:
* Write a short, friendly tweet about it.
* Create a 3-paragraph blog post summarizing the news.

Save this file as instructions.md.

Step 2: Hand the file to your AI Agent

You can upload this Markdown file to an AI tool (like ChatGPT, Claude, or a custom AI Agent). Because you used Markdown, the AI immediately understands which part is the title (#) and which parts are the tasks (*).

Step 3: Let the AI Agent do the work!

The AI Agent takes your “recipe” (the Markdown file), reads it, thinks about it, goes out to find the space news, and writes your tweet and blog post.

The cool thing is you can take your recipe to any AI LLMs!

Last

You did it! You now know the exact difference between a Markdown file and an AI Agent. You’ve learned that one is a simple way to format text, and the other is a powerful digital assistant, and that they make a fantastic team.

Your next step: Try writing a tiny bit of Markdown yourself! Open up a free AI tool like ChatGPT, and type a prompt using Markdown symbols, like # My Favorite Hobbies followed by a bulleted list using asterisks *. Watch how the AI easily reads your beautifully organised text!

Happy exploring!

Phil

Imagine you have a super‑helpful student teacher in your classroom.
You can say, “Can you create a quiz from yesterday’s lesson, email it to the parents, and then put the marks into the gradebook?”
They don’t just answer your question and smile; they actually go off, do the work, and come back when it is done.

That is roughly what an “agent” is in AI language, in teacher terms.

The simple definition

You can explain it like this:

“An AI agent is a digital helper that can do tasks for you, not just answer questions.
You give it a goal, and it thinks through the steps, uses tools (like email, calendar, or documents), and carries out the work on your behalf.”

Key bits, in plain language:

• It works on your behalf (like a helper).

• It can decide the steps to take, instead of you micromanaging every click.

• It can use tools, not just chat – for example, it might read a document, create another one, or send a message.

One extended analogy: the student teacher

Think of an AI agent as a very capable student teacher in your classroom:

• You say: “Please create a worksheet on fractions, based on last week’s notes, and put it in Google Drive.”

• The student teacher:

  • Reads your notes.

  • Designs questions at the right level.

  • Formats them nicely.

  • Saves them in the right folder.

With an AI agent:

• You write: “Create a worksheet on equivalent fractions for Year 6 and save it to my drive and <here’s my full context getting you up to speed and hoping that you will pick it up or I will keep adding to this context window>.”

• The agent:

  • Reads your curriculum document or previous materials.

  • Is guided by the context.

  • Writes suitable questions.

  • Uses connected tools (like Google Drive, email, calendar) to put that work where it needs to go.

A normal AI chatbot is like a smart student who talks with you but stays in their chair.
An AI agent is like a student teacher who gets up, walks around, and actually does things in the school.

What’s happening “behind the scenes” (without the jargon)

Under the hood, most agents do a loop a bit like this:

• Understand your goal and context better with memory or databases.
  Turn your request (“make a quiz and email it”) into a clear set of tasks.

• Plan steps
  Decide on a rough plan: “First create questions, then format them, then send the email.”

• Use tools
  Call other apps or services: your email, calendar, documents, learning platform.

• Check their own work
  Look at the result and adjust: “Are these questions at the right level? Did the email send?”

You don’t need to explain the technical details.
For a non‑tech teacher, it is enough to say: “It understands what you want, figures out the steps, and uses other apps to get it done.”

A 3‑step way you might explain it to a colleague

You could walk another teacher through this:

1. Start with what they know
   “You know how ChatGPT or another AI will answer your questions in a chat?”
   “That’s like having a very smart student to talk to.”

2. Add the ‘does things for you’ idea
   “An agent is the next level. Instead of just talking, it can actually act – create files, send messages, set reminders, organise information.”

3. Use a school‑based example
   “For example, you could say:

   • ‘Create a parent email explaining our upcoming exam and draft it in my Gmail.’

   • ‘Read this lesson plan, then generate a 10‑question exit ticket in a Google Doc.’
     The agent figures out the steps and does the boring parts for you.”

If they seem nervous, you can add:
“You still stay in control. You can always check and approve what it does, just like a mentor teacher checks a student teacher’s work. Lead with your context, and be open to re-design parts of your learning and assessment expeience.”

Why this matters for a teacher

For a teacher with no tech background, the important message is not the buzzword “agent.”
The important message is: “Soon, some AI tools won’t just give you ideas; they will help you actually do the admin work that eats your time. They will connect to your context so you won’t always have to say the full speel.”

For example, agents can help with:

• Drafting and filing lesson resources.

• Generating differentiated versions of the same task.

• Summarising student work or behaviour notes.

• Preparing parent communications from your rough notes.

So when they hear “agent,” they can simply think:
“Ah, that’s the version of AI that doesn’t just talk to me; it also does tasks for me.”

Phil

Have you ever written an email, read it back to yourself, and thought, “Yikes, that sounds a little too grumpy”? So, you delete a few words, add a smiley face, and rewrite it before hitting send.

Believe it or not, Artificial Intelligence can do exactly the same thing if you ask it!

As beginners, we often think that whatever the AI gives us on its first try is its “final answer.” But the real magic happens when you ask the AI to put on its editing hat and critique its own work.

What Does “Self-Critique” Mean for AI?

Think of an AI like an eager, lightning-fast sous-chef. If you ask for a bowl of tomato soup, it will quickly chop, boil, and hand you a bowl. But what if you stop and say, “Taste this soup and tell me what it’s missing”?

The chef might pause, take a sip, and realise, “Ah, it needs a pinch of salt and a little more simmer time.”

When you ask an AI to critique itself, you are simply forcing it to pause, review its own generated text against your original instructions, and find its own mistakes. It’s a strange loop but it gets better results. I use a critique-yourself skill on Claude all the time.

The “Magic” Behind the Mirror

When you say, “critique your previous answer,” a few wonderful things happen. First off, AI doesn’t have an ego to protect. It won’t get defensive or have its feelings hurt! Instead, it smoothly switches gears from “creative mode” to “analytical mode.”

Here is what the AI is actually doing behind the scenes:

• It checks the rules: Did it follow all of your instructions? Did it make the text as short or as long as you wanted?

• It looks for gaps: Did it forget something important?

• It improves clarity: Is the language too complicated? Does it sound like a robot wrote it?

By asking it to look for these flaws, the AI spots its own weaknesses so you don’t have to.

How to Make AI Critique Itself (Step-by-Step)

Ready to try it out? It’s incredibly easy and requires zero technical skills. You can use this simple 3-step process on ChatGPT, Claude, or any other friendly AI you like to use.

Step 1: The Initial Ask

Start by asking the AI to do a task, just like normal.

• What you type: “Write a short, friendly email to my neighbors inviting them to a summer backyard BBQ.”

Step 2: The Self-Critique

Once the AI gives you the first draft of the email, don’t copy and paste it just yet. Instead, ask it to review its work.

• What you type: “Critique the email you just wrote. What could be improved? Are there any details missing, or is the tone too formal?”

• What happens: The AI will reply with a helpful list of its own flaws! It might say something like: “The tone is a bit stiff for a neighborhood party. Also, I forgot to include placeholder blanks for the date, time, and address!”

Step 3: The Rewrite

Now, tell the AI to use its own feedback to make a better version.

• What you type: “Great points! Please rewrite the email based on your own critique. Show me how you improved it!”

• What happens: You get a second, much better draft that feels more natural, friendly, and complete.

Why This is a Superpower for Beginners

When you’re just starting out with AI, you might not know exactly how to write the perfect instruction (often called a “prompt”). And that is perfectly okay!

By using the self-critique method, you don’t have to be a master prompt writer. You and the AI become a team. You provide the spark of an idea, and you let the AI do the heavy lifting of polishing it until it sparkles.

Your Next Steps

Next time you use AI to help you write a message, brainstorm ideas, or learn a new topic, don’t settle for the very first draft. Ask the chef to taste the soup! Ask the AI to grade its own homework. And review the reasoning it gives.

Phil

Why does telling an AI it is a “world-class expert” often result in bland, cliché-heavy “LLM slop” rather than actual high-level insight?

We will examine how persona prompting triggers statistically “safe” but generic training data. Then, we’ll provide a workflow to replace empty titles with structural constraints that actually preserve a human voice.

The Architecture

The following system instruction forces the model to ignore its “expert assistant” defaults and prioritise specific, idiosyncratic human writing patterns over generic polish.

As an editor, prioritize "human-readability" over "machine-perfection." Do not use stock phrases like "In today's world" or "At its core." 
When I provide a draft, correct only factual errors and confusing grammar. 
Do not rearrange sentence order, do not add "In conclusion" summaries, and do not soften my opinions with hedges like "it's important to consider." 
If a sentence is punchy but technically a fragment, leave it. Focus on specific nouns and active verbs.

The Workflow

Step 1 - The human must define the “Spiky Point of View.” Before touching the AI, write one sentence that is controversial or specific to your experience that a generic “expert” would never say.

Step 2 - Use a “Scoped Editing” prompt instead of a persona. Instead of “Write like an expert,” use: “Fix the grammar in this voice dump but do not change my word choice or sentence rhythm. Preserve the ‘scuffs’ in the prose.”

Step 3 - Force the model to show its work using a “Diff.” If you suspect the AI is “blandifying” your work, use this instruction: “Rewrite this paragraph for clarity, but bold every word you changed so I can see where you’ve sanitised my voice.”

Developed by Phillip Alcock, Co-Founder PBL Future Labs

About the Author:

Phillip Alcock is an AI in education strategist, consultant, and curriculum designer. He focuses on building practical, zero-BS workflows that help educators and professionals actually use AI without losing their minds or their voice. Find more frameworks at PhillipAlcock.com.

How do you stop an AI from ignoring half your custom instructions when you hand it a complex, multi-step task?

The Approach

You stop dumping a massive wall of text into the prompt window and start treating the AI like a literal router. I normally setup projects. By using a central plan.md file to sequence your separate sub-files, you force the model to process one narrow constraint at a time.

The Architecture

Put something like this into a Google Gem/Claude Project:

I am assigning you a multi-step writing task. Before you generate any output, read `plan.md`.

`plan.md` contains the exact sequence of steps you must follow as the orchestrator.
It will explicitly tell you when to reference `structure.md` for formatting rules, and when to reference `voice.md` for style constraints.

Do not skip ahead. Do not execute Step 2 until Step 1 is entirely complete. 

The Workflow

Step 1 - The first step is always human. You have to map out your dependencies before you hit enter. Figure out exactly what the AI needs to know for structure, what it needs for style, and drop those into separate markdown files. If you just dump everything into one mega-prompt without a plan, you are just running the slot machine.

Step 2 - Put it into the AI with these instructions to establish the routing rules:

Read the attached files. Follow the sequence in `plan.md` exactly.
Step 1: Build the outline using only the constraints in `structure.md`. Stop and wait for my approval.
Step 2: Ask me 10 questions about the article to ensure I understand it.

Step 3 - Extend and re-iterate using these instructions when the AI inevitably loses focus:

The draft drifted off track in paragraph three. Look back at Step 2 of `plan.md`. Re-apply the constraints from 'workflowprocess.md'

Developed by Phillip Alcock, Co-Founder PBL Future Labs

About the Author:

Phillip Alcock is an AI in education strategist, consultant, and curriculum designer. He focuses on building practical, zero-BS workflows that help educators and professionals actually use AI without losing their minds or their voice. Find more frameworks at PhillipAlcock.com.

Before: You open Claude.ai and describe your PBL unit. The driving question, the timeline, the student groups, the milestones you need to track. You spend ten minutes typing context you already have written somewhere else. You get a decent worksheet. You close the tab. Three weeks later you need to update it for a new class. You open a new chat and type everything again.

After: You have a folder on your computer with one brief file inside it. The brief describes the unit once, clearly, in plain language. You open Cowork and point it at the folder. It reads the brief. It produces the worksheet. When the unit changes, you update the brief. You run Cowork again. The new worksheet reflects the change. You never describe the unit from scratch again.

Where the context lives is the whole difference.

Here is the idea, as simply as I can put it.

Claude.ai is a chatbot. When you open a new conversation, it knows nothing about you except what you type in that message. Every conversation starts from zero. That is not a flaw it is how chatbots work. You bring the context every time, in the message.

Cowork is something different. It is a desktop agent, which means it works the other way around. Instead of you building context in a message, Cowork reads context from files on your computer. Instead of you waiting for a response, it runs a plan a sequence of steps you write once and reuse. When the session ends, the files stay. The context persists. The next time you run it, nothing has been forgotten.

Let me say that another way.

Think about the difference between phoning a knowledgeable stranger and briefing a colleague. The stranger is helpful but you have to start from the beginning every time , what your project is, who your students are, what format you need, why the timeline is what it is. The colleague already has the folder. You say "can you update the tracking sheet for Week 4" and they know exactly what you mean, because the brief is sitting on the desk between you.

Here is the simplest version.

A chatbot answers. An agent works.

That is the whole trick.

The Workflow

The fastest way to feel the difference is to try the same task both ways.

Start by opening Claude.ai and describing a PBL unit you are currently running. Ask it to produce a student tracking worksheet. You will get something useful. Now notice what it cost you, several minutes typing context you already have written somewhere else, and notice what happens when you close the tab: that context is gone. Tomorrow, if you want a variation, you start again.

Now try it the other way. Open a text editor and write your unit down in plain language, the driving question, the student groups, the timeline, what you need to track. Save it as a file called brief.md in a folder on your desktop. This takes the same time as typing the context into a chatbot. The difference is that you only do it once.

In the same folder, create a second file. Call it plan.md. Write one sentence: "Read brief.md and produce a student tracking worksheet as a table." That is your plan. It does not need to be long. It needs to be clear.

Open Cowork and select that folder. Cowork reads both files. It knows what the unit is and what you need. It runs the plan.

Now go back to brief.md and change one thing, a date, a group size, a milestone. Open Cowork again. The new output reflects the change. You did not re-explain anything. You updated a source file, and the output updated with it.

This is the shift. When context lives in a file rather than in a chat message, every future output inherits that context automatically. You are not talking to an assistant each time. You are briefing one, once, and the brief stays current because you maintain it.

The chatbot was useful. The agent is reliable. Those are different things, and once you feel the difference you will not go back.

The Prompt Lab (For subscribers)

Five prompts to get your first Cowork folder running this week.

The minimal brief

Write this in a new file called brief.md in any folder on your computer. Fill in the details for a real unit or project you are running now.

"Unit title: [your unit]. Driving question: [your driving question]. Duration: [X weeks]. Number of student groups: [X]. Group size: [X]. Milestones: [three key assessable moments you want to track]. Format preference: table."

This is enough to start. It does not have to be comprehensive it has to be clear. I often just set minimal constaints, and add constraints as I need.

The minimal plan

Write this in a second file called plan.md in the same folder.

"Read brief.md. Using the unit details provided, produce a student group tracking worksheet. Format it as a table with group names across the top and milestones down the left side. Include a notes column for each group at the end."

Run this with Cowork and look at what comes back. Then open brief.md and change one thing. Run it again. Compare the two outputs.

The context upgrade

Once your first version is working, add a second section to brief.md.

"My context: [brief description of your students — age, school type, any relevant details]. My constraints: [time available, tools you have, what has worked before]. My format preference: [e.g. 'I print these, so plain text tables work better than formatted HTML']."

Every output will now reflect this context without you repeating it in the plan.

The revision test

After your first successful run, deliberately break something. Change the timeline in brief.md to something incorrect. Run the plan again. Notice how the output changes. Then fix the brief. This is the exercise: the brief drives the output, so the brief has to stay accurate. Update the source and the output corrects itself.

The delegation question

Before writing your next brief, ask yourself: "If I handed this folder to a capable colleague who had never heard of this project, would they have everything they needed to do the task?" If yes, the brief is ready. If not, add what is missing. This is the standard a brief file needs to meet, not for Cowork, for you.

Good luck!

Phil