AI Postcards is a StrefaTECH series of short notes about my journey with AI. Each shares one thing I actually did with AI lately, and something it taught me. Hopefully, you’ll find these to be quick reads that share some highlights from my AI travels!

I recently joined a new community choir and received a printed schedule for the season, including the rehearsal schedule with locations, times, and various other notes (e.g., bolded list of mandatory sessions!).

When I got home, I considered sitting down to add them to my calendar one by one. But instead, I took two photos of the printed pages, dropped them into AI1, and typed: “Give me this in text.” What came back was exactly that: a clean text version of the schedule, dates, times, locations, all correctly pulled off the printed page.

Then I asked for what I really wanted. I next told it that I wanted all of the sessions on my calendar with an event before each for an hour of transit. One clarifying question later (I have two calendars — which one?), and it was done.

What came back was two calendar entries for every session: one for the rehearsal or performance, location included, and one for the transit block right before it.

If you’d handed me those same two pages a year ago, I’d have opened my calendar and just started clicking and typing; it wouldn’t have occurred to me there was another way. I keep noticing lately that for many tasks, I technically could do all of the work myself, but if it’s not a simple one-off, having AI do it works nicely.

The saga continues … some calendar cleanup

A week later, our choir’s assistant director emailed asking which events I could make. (I had mentioned some fall conflicts when I auditioned.) I wrote back a plain yes-and-no list2. Then I pasted that same reply into the chat with Claude with the simple outcome I wanted: “Fix my calendar.”

Before touching anything, it laid out which events it planned to remove or edit and asked for a single go-ahead. Once I approved, it handled the whole batch. It also flagged two things I hadn’t caught myself: a scheduling conflict later in the fall I still need to sort out, and a pair of performances my email hadn’t addressed at all, so it couldn’t tell whether I was doing them.

This is the second time this pattern has shown up for me lately (Issue #157 was the first): outcome in, reasoning handled. But handled doesn't mean flawless.

Hiccups and hallucinations are still worth checking for

Sometimes, what comes back is exactly what I expected. Other times, though, it’s wrong in a way I have to catch.

For example, when I later asked for a reminder on those transit blocks and asked it to double-check that some old reminders were really gone, it told me everything was clean. It wasn’t; one was still firing. A screenshot later, it found a bug in how the calendar tool handles “clear this” requests, and admitted that testing the fix had briefly made one event worse before it landed on the actual solution. (Fortunately it noticed and told me how to fix its goof.) It also explained to me a nuance of how Google Calendar default reminders work; learn something new every day!

The bottom line

So how much of this was actually “conversation”?

Claude asked one question about which calendar to use and for one approval before any batch of edits went through. I gave it a screenshot proving something it had called “clean” wasn’t. That’s not nothing, but it’s also not me doing the work.

Every one of those exchanges was a decision only I could make, or a mistake only I could catch. I needed to remain engaged in the process, delegating the grunt work to AI and saving myself a bunch of time and clicking in return.

In short, aim it at an outcome and let it figure out the how, but don’t wander off while it works!

And as you’re looking at your own overflowing calendar this season, think about where and how AI can help, and as always...

AI Snips and Tidbits is a StrefaTECH grab bag of quick tips from my AI travels. No big lessons here, just small things I did with AI recently that worked (or surprised me!), shared while they’re fresh. Quick reads, no schedule, hopefully a nugget or two you can use today!

Run your research twice

I gave the same research task to Claude’s Research and ChatGPT’s Deep Research. They both gave me the same big conclusion, but each missed about half of what the other found. One dug through institutional documents while the other reasoned about people and context. I was glad I asked twice... there were truly good insights from both!

The tip: For anything consequential, run it in two tools (if you can) and compare. Want to try research without paying? ChatGPT and Gemini include a small monthly allowance of deep research (as of this writing); Claude’s requires a paid plan.

Spend time prepping for Deep Research

Speaking of deep research, here’s how I set one up this week. I’m evaluating conference room equipment (microphones, video, speakers) for a nonprofit board I serve on, so that our hybrid in-person/Zoom meetings will work better for everyone. Before launching ChatGPT’s Deep Research, I described the situation and asked it to write me a detailed research prompt. We went back and forth on a couple of questions it raised, and the final prompt included considerations I would have missed entirely if I’d just sent it off to research on its own.

The tip: Use AI as a brainstorming partner before just giving it some assignment. It can be quick and result in some good insights! And as in so many cases, talking to the computer (e.g., with WisprFlow) is a huge win.

Summarize responses to emails (lazy but oh-so-nice!)

I’m Treasurer of a nonprofit board, and I needed to schedule recurring Finance Committee meetings for the rest of the fiscal year. This wasn’t a job for Doodle or When2Meet (my usual meeting-scheduling-outreach go-to’s), and I dreaded the usual routine... everyone replies to my email in a different format, and I scribble it all on a notepad with hopes of finding a decent slot.

This time I asked AI to summarize the responses sitting in my email and recommend a time. I gave the task to Claude, ChatGPT, and Gemini just to see what I’d get, and all three did it far better than my notepad ever has! Claude edged out the others (while also taking the longest to do it), and it went on to draft the follow-up email and calendar invitations.

The tip: Connecting your AI chatbot to your email and calendar can be hugely helpful. A warning, though: be very careful to instruct it to draft only, never send.

Tell the AI your outcome, not your solution

I needed to collect photos from a dozen or so family members with wildly different tech comfort levels for an upcoming family gathering. I could have sent an email or built a Google Form. Instead, I described the outcome I wanted (easy for everyone, minimal work for me) to Claude and let it propose options and ask me questions back. That conversation surfaced a gotcha I would have hit later: my first-choice tool made collecting the photos easy but downloading and using them painfully clunky. Starting from the goal instead of a tool got me to a better answer than I’d have picked on my own. (I wrote about this idea in issue #146 back in March... which in AI time might as well be a different century!)

The tip: Tell AI as much as you can about why you want to do something, have a conversation, and then let it rip!

That empties the tip log for now, though it never stays empty for long. Have a small AI win of your own? Share it in the Comments or drop me a note at strefatech.info@gmail.com. And while you’re gathering your own snips and tidbits...

AI Snips and Tidbits is a StrefaTECH grab-bag of quick tips from my AI travels. No big lessons here, just small things I did with AI recently that worked (or surprised me!), shared while they’re fresh. Quick reads, no schedule, hopefully a nugget or two you can use today!

Welcome to the very first batch of AI Snips & Tidbits.1 Today I’m sharing four things I did with AI lately that stuck with me and might be a small inspiration for you.

Helping a family member move into a new house, I had a little plastic doohickey in my hand and wasn’t positive it was what I needed. So I snapped a photo with my phone and asked ChatGPT, which told me it was a wall anchor and started walking me through what to do with it. I hammered it partway in, took another photo, and got back “not quite, needs to go farther.” Tapped it flush, one more photo, and this time a (virtual) thumbs up!

I do variations of this sequence (photo, question, action, another photo) so often now that it’s becoming a habit ... and 9 times out of 10, I’m delighted to have the help.

Claude’s drafts of my articles kept sounding like AI, even after we’d fixed the tells one by one. So I flipped it around. I uploaded a batch of my old articles from before I ever touched AI, and instead of asking what to avoid, I asked what my writing always does: “You are an expert in writing human-sounding articles. Critique this draft and list your suggested changes to make it sound more like me writing it myself. Attached are examples of my early, non-AI-assisted articles.” It named things I never would have (apparently I can’t write three sentences without a parenthetical, which, well ... guilty!). The trick turned out to be giving AI examples of the real you, not just a list of don’ts.

Four of us were on the porch one evening, talking about insects the way you do when nobody wants to go inside yet. Somebody wondered what a fishfly eats, so I asked ChatGPT out loud on my phone ... and it just stayed in the conversation with us. Fishflies to dobsonflies to caddisflies, everybody talking over everybody and doubling back (“no, no, what does a dobsonfly eat?”), and it kept up with every one of us. It even pulled up photos when we couldn’t picture what we were describing. I’ve handed AI plenty of solo tasks, but keeping pace with a rambling four-way porch chat? Did not see that coming!

At the end of a marathon session with Claude, where we’d worked out a pile of new writing rules and a whole workflow, it hit me that all of it would vanish the second I closed the chat. So before I did: “Give me updated complete project custom instructions so I can just replace what’s there.” I pasted the result into my project settings, and now every new chat starts already knowing what took us weeks to figure out. I’m becoming a big fan of reviewing my custom instructions now and again, and having the AI app help me to improve them.

That’s the debut. Four snips, no schedule, and I’ll share more whenever a few good ones pile up.

So how about you ... did one of these spark something you'd try yourself? Tell me in the comments (or at strefatech.info@gmail.com). And as you go poking at what AI can do in your own corners of the day ...

AI Postcards is a StrefaTECH series of short notes about my journey with AI. Each shares one thing I actually did with AI lately, and something it taught me. Hopefully, you’ll find these to be quick reads that share some highlights from my AI travels!

My mother-in-law’s birthday celebration is coming up, and family is flying in from all over, which means I need a way to collect travel details and a few other pieces of information from everyone before they descend on our weekend Poconos retreat at the same time.

First, I had a conversation with ChatGPT. I described who was coming, what I needed to know, and why. Then I asked it to work with me on a prompt I could hand to a form builder. We went back and forth on phrasing and what kinds of options to include, and I ended up with something much clearer than what I would have typed on my own. (The conversation also helped me figure out what I actually needed to know, which I hadn’t fully thought through!)

I pasted that prompt into Google Forms, into a feature I hadn’t used before called “Help me create a form.” It builds the whole thing from a text description, using Gemini on the back end. The form appeared in about two seconds (I’m not exaggerating!), and it was almost exactly what I wanted, except one field I needed marked required wasn’t. Two clicks to fix that, and very impressive overall.

That’s not actually the whole story, though. Before I ever found that button, I’d already built a working version of the same form myself, the hard way, using AI-written Apps Script code. It mostly worked ... until I hit an error message that meant nothing to me. I took a screenshot, pasted it into ChatGPT, and it told me this was a fairly common Apps Script oddity and to try running it again in an incognito window. That actually worked, but for a minute there, I was ready to just bag the whole thing and build the darned form by hand.

So somewhere in my Google account right now, there’s a form built by code I couldn’t have written myself, solving a problem I could have solved in two seconds if I’d known where to look. (I’m keeping it, out of pure stubbornness.)

The moral of this story?

AI features keep showing up inside ordinary apps faster than I can keep up with!

This one had been sitting inside Google Forms the whole time, and I had no idea until I stumbled onto it by accident.

Getting ChatGPT to create the prompt for Google Form’s helper is basically the same trick from issue #155, where I found out you can just ask AI to write your image generation prompt instead of trying to craft it yourself. Have the conversation first, then hand the result off to whatever tool needs it. It doesn’t matter whether that tool is a chatbot’s image generator or a form builder’s “help me create” button. Knowing the button’s there in the first place is the harder part.

I don’t have a system for catching these ahead of time, and honestly, I’ll probably build another workaround before I notice the next one. But next time, before you build a workaround (or ask AI to build one for you), maybe take a minute to poke around the app you’re already in. You might find that the feature you need was added while you weren’t looking.1

Have you stumbled onto an AI feature hiding in an app you use every day? Share your tip in the comments!

And as you go looking...

1

While I had ChatGPT open, I also asked it to generate a banner image for the form using a photo of my mother-in-law and asking for some travel and birthday themes. Here’s a ‘sanitized’ version that ChatGPT created when I told it I needed something similar for this article that didn’t actually use her photo!

AI Postcards is a StrefaTECH series of short notes about my journey with AI. Each shares one thing I actually did with AI lately, and something it taught me. Hopefully, you’ll find these to be quick reads that share some highlights from my AI travels!

I almost always ask the AI to ask me questions before I ask it to do anything.

As I was working with Claude to develop a new article series (the one you’re reading right now), instead of just telling it what I wanted, I asked it to interview me first. What came back was five or six questions. I answered maybe two-thirds of them. The rest I skipped, either because they didn’t apply or because I didn’t have a good answer. (Claude filled in the blanks.) Then we got to work.

I wrote about this technique briefly in Issue #146, but I’ve been sitting with why it actually works, and I have more to say about it now.

The obvious answer is that the questions fill in gaps. The AI surfaces things it needs to know that I forgot to include, and the response is better for it. That’s real, but it’s not the interesting part.

What I didn’t expect is what answering the questions does to me.

When I’m just prompting and waiting, I’m offloading—I describe what I want, hand it off, and see what comes back. That works fine for a lot of things. But answering questions is different. I’m still in the conversation. And something happens when I’m forced to articulate things out loud (yes, I talk to my computer, as you may recall!1) that just doesn’t happen when I’m passively waiting for output. Thoughts I didn’t know I had start to show up.

The AI isn’t doing that. I am.

Cognitive scientists have been studying this pattern—what happens to our thinking when we offload too much—and the short version is that the mental work we skip is also the mental work that builds understanding.2 The AI created the conditions for me to find an insight I wouldn’t have found on my own. That feels meaningfully different from the AI just handing me an answer, even if the end result looks similar.

Try it with whatever AI you use—Claude, ChatGPT, Gemini, it works the same way in all of them. Describe what you’re trying to do, then add a line asking it to interview you before it responds. Answer what’s useful, skip what isn’t. See if anything unexpected comes up.

And as you do...

I’m a member of the Bloomsburg Singers, a community choir. Our director and choir managers send out weekly emails with rehearsal schedules, links to online recordings, repertoire updates, logistics. I’ve been meaning to file them into a labeled Gmail folder and, well, some of them made it there and some didn’t. (My inbox is not a model of organization recently!)

With our concert coming up in a few weeks, I decided I wanted to get myself organized, and it occurred to me that I could ask AI to help me pull everything together. I realized that creating a simple webpage was likely to be about as easy as putting the information anywhere else, with the bonus that I could share it with the whole choir with a single link.

What's "vibe coding," anyway?

You may have seen the term “vibe coding” floating around. It has the word “coding” in it, which can be intimidating (and these days has little to do with what we do with these apps from our perspective), so I’m going to call it “AI-assisted building” instead. The idea is that you describe what you want in plain language and the AI handles the technical side, without you ever writing (or seeing!) a line of code.

Tools in this space include Claude Code, OpenAI Codex, Lovable, Bolt.new, Replit, and Cursor, among others. Most offer some limited free use, though the limits can get annoying fast. If you’re already paying for an AI subscription, it’s worth checking whether building tools are included (e.g., Claude Code comes with Claude’s paid plan, and Codex comes with ChatGPT).

How did my webpage creation go?

I didn’t come in with a tidy list of specs.

I gave Claude a rambling description of what I had in mind and then asked it to interview me to flesh out the details.1

It did, asking quite a few useful questions about design, what contact information to include, a few ambiguous pieces of info from the emails, and where I wanted to host the page2. That interview-style kickoff is something I’d suggest to anyone who isn’t sure how to start a prompt. It works well.

Claude then pulled eleven emails from Gmail3 (some from the Bloomsburg Singers label, some still sitting in my mess of an inbox), extracted the concert repertoire, YouTube links, rehearsal logistics, and contact information, and built a clean, mobile-friendly page. For the color scheme, I pasted in an image of our concert poster^, and Claude pulled the navy, red, and sky blue right from it. My fairly simple instructions got a crisp page design that worked well for something that will only live for a few weeks. With more time and more detail in the prompts, I could have pushed it further toward a stronger brand look, but for this purpose, it was exactly right.

As a bonus, I asked Claude whether it could create a sequential YouTube playlist of all our pieces in performance order. It could, and it was simpler than I expected. I used the playlist myself while making dinner, running through the repertoire while I cooked.4

Near the end of the session, I asked Claude to draft an email to our director and choir manager about the new page. It took maybe thirty seconds because everything needed was already in the conversation: who they are, what’s on the site, and what was new. When you’re working on a longer project with one of these tools, that accumulated context becomes genuinely useful.

The whole session ran about an hour and a half, though my actual attention during that time was probably 30 to 40 minutes, since I was multitasking quite a bit while it worked. And it was a much more enjoyable experience than would have been the case had I needed to dig through my emails to find all the relevant information and copy/paste into a Google Doc. It might not have been faster in the end, but the result was way better and much less tedious!

What did I catch before I hit publish?

Before I finalized the page, I did what I try to do with anything AI helps me create: I stepped away, came back with fresh eyes, and thought about what it actually contains, remembering that I’m on the hook for anything that’s wrong or risky. This page has no password protection and nothing stopping search engines or bots from finding it, which means anything on it is genuinely public. I had included contact emails for several people in the choir, and I wasn’t comfortable leaving personal email addresses exposed that way without their permission. So I asked Claude Code to remove the email addresses.5

A few minutes of that kind of check is worth building into your process. Hallucinations, information you didn’t mean to expose, details that looked right but weren’t … all of those are things that fresh eyes can catch and that you might miss when you’re deep in a project. (You can ask AI to do a check, too … but for me, that final step of “approving” that something’s reading to send/post/publish is on me to do.)

Could this work for you?

A choir resource page is one flavor of what AI-assisted building can produce. Event pages, a simple landing page for a fundraiser or program, a reference page for a volunteer team, a one-page site for a small organization that doesn’t have one yet -- these are all well within reach once you get comfortable with the back-and-forth of describing what you want and refining from there.

The more you use it, the more you start noticing information you want to share with others and thinking:

I could just make a page for this.

Once that instinct kicks in, it’s a pretty short trip to actually doing it, and when you do...

My friend Tim has had quite a life. Activist, musician, and inspiration to a lot of people who know him. When his birthday came around recently, I wanted to do something more personal than a card from the store or text message with a stock gif. So I pulled a handful of photos of him from different stages of his life, uploaded them to ChatGPT, and asked it to turn them into a birthday collage. What you see above is what came back.

That got me going. Mother’s Day was coming up, and I have lots of pictures of my mom and mother-in-law, the kind that just waiting in Google Photos for me to do something with them. This year I did! I uploaded a few to ChatGPT, described what I wanted, and within minutes I had a dozen personalized Mother’s Day images: watercolor-style, spring flowers, a banner at the bottom with messages ChatGPT suggested on its own, things like “I love you, Mom” and “Here’s to you” alongside “Happy Mother’s Day.” I loaded them onto the digital photo frames my mom and mother-in-law both have.1

Next was my daughter’s wedding anniversary; five minutes with ChatGPT and my photo library, and I had a pretty card image to text. Two weeks of this, and I’m genuinely hooked!

Here’s how it works.

The Part That’s Easy To Skip: Ask AI to Write the Prompt

The most useful thing I can tell you is this: you don’t need to figure out the prompt yourself. Just tell ChatGPT what you want in plain English, and ask it to write the image generation prompt for you.

I described what I was going for (dictating with WisprFlow as usual!) and ChatGPT handed me a detailed, carefully worded prompt ready to use. I didn’t have to think about things like “soft pastel washes” or “gentle hand-painted texture.” It handled the vocabulary of image generation so I didn’t have to learn it.

If you’ve read my earlier piece on conversation engineering (issue #132), this is exactly that idea in action. You’re not programming the machine; you’re having a conversation with it. Start with what you want, and let it work out the technical language.

The Basic Process

Once you have your prompt, the rest is pretty straightforward. Upload your photo (or photos; I used seven for the birthday collage) to ChatGPT2, paste in your prompt, and hit generate. You’ll get an image back. If something’s off, you can tell it what to fix in plain language, or use the Edit tool to select a specific area and describe the problem (e.g., ”the face in the upper left doesn’t look right”), and it will try again.

Here’s the prompt I used for the Mother’s Day images:

Transform this photo into a warm, colorful Mother’s Day greeting-card style image for a digital photo frame. Keep the people recognizable and preserve their facial features, expressions, body positions, and relationships as closely as possible. Do not change who is in the photo. Keep the overall composition similar, but make the image feel more cheerful, polished, and artistic. Style: bright watercolor illustration with soft pastel washes, gentle hand-painted texture, warm lighting, and a joyful spring atmosphere. Add decorative spring elements around the edges: flowers, tulips, daisies, butterflies, small hearts, greenery, and a smiling sunshine. Keep the decorations framing the photo, not covering faces. Add a large decorative ribbon banner across the bottom. The banner should look hand-painted, cream or pale yellow, with soft pink/purple lettering. The banner should have a message like: “I love you, Mom” “Happy Mother’s Day” “You’re the best!” Make the text large, clear, correctly spelled, and easy to read. Do not add any other words. Format: horizontal landscape image, suitable for a digital photo frame, high resolution, cheerful, affectionate, family-friendly.

For the birthday collage, I uploaded seven different photos of my friend Tim at various stages of his life and asked for something similar: watercolor background, a personalized message, decorative elements.

When It Doesn’t Quite Work

A word of warning: it’s not perfect every time. In one of the family photos, one person’s face just looked slightly off across three separate attempts, very close but not quite right. After the third try, I moved on. Regenerating indefinitely gets diminishing returns, and honestly, most of the images I got were good enough to send with a big smile.

The faces that are hardest for the AI are usually ones where the source photo isn’t great, small, slightly blurry, or partially turned. A better source photo generally means a better result. Worth knowing too: the AI can sometimes nail a person’s likeness on the first try and then be noticeably off on the next attempt with the same photo. There’s no predicting it, so you just look at each image as it comes.

Other Things You Could Do With This

Cards and photo frames are the obvious use cases, but once you start playing with it, you’ll think of others, like birthday posters, retirement tributes, and holiday images for a family group chat. If you’re in an organization, think about event programs, volunteer appreciation visuals, or donor recognition images … anywhere a personalized touch would land better than a stock photo.

A Note on Ethics and the Environment

A few things worth thinking about before you dive in.

The photos I used for personal cards are of real people, and I kept those images private, shared only with the people in them or others I knew would be OK. The birthday collage on this article’s cover is different; those were images already publicly available online, so I felt comfortable using them here. I’ve thought about where that line is, and I think you should too.

Transformed or not, a photo of someone is still a photo of someone.

On the environmental side: image generation uses more computing power than a text prompt, which means a slightly higher carbon footprint. I’ve made my own peace with the tradeoff for creating these images. I’m generating them instead of buying cards that get manufactured, shipped, and eventually thrown away. But I also stopped after three attempts on the image that wasn’t working. That felt like the right call, and not just for the environment. At some point you’re just spinning your wheels.

And, as I’ve said before in this newsletter, each of us gets to decide how we use these tools and how much is enough. There’s no universal answer, but it’s worth asking the question. Which, come to think of it, applies to pretty much every AI decision you’ll make, so remember to consciously...

“My AI Confessions” is an occasional series where I come clean about the tools I use, the approaches I take, and whether they’re actually working.

I gave an AI tool a simple task: pull an audio file from my Google Drive, transcribe it, and save the transcript back to that same folder. The kind of thing I ask AI tools to do constantly now. It came back almost immediately, confirmed it had processed the audio, and gave me a link to the Google Doc with the transcript.

I didn’t even open it. The tool said it was done, so as far as I was concerned, it was done.

It wasn’t until I started asking questions about the conversation in that recording -- I was a participant, so I had some idea what was actually said -- that the responses started getting strange. Not a little off, they were completely nonsensical! When I pushed back, the tool apologized and explained that it had fabricated the entire transcript. It couldn’t access audio from a Google Drive link. The document it had pointed me to was fiction, and I’d never even looked at it.

I was genuinely angry. And then, once the anger had settled, honestly I was kind of embarrassed. What finally tipped me off wasn’t careful verification on my part -- it was that the hallucinations were so outrageous I couldn’t keep rationalizing them. A more plausible fabrication might have slipped right by me.

That’s the whopper that woke me up.

The Danger Zone, Revisited

About a year ago, I wrote about what I called the AI Danger Zone.1 The concept borrowed from learning to drive: new drivers are terrified and hypervigilant, mirrors checked obsessively, both hands on the wheel. Then familiarity sets in. The music goes up, the phone comes out, and what used to require full attention becomes autopilot.

With AI tools, the Danger Zone is that stretch when you’re comfortable enough to trust the output but not careful enough to verify it. You stop treating responses as drafts to be checked and start treating them as answers.

What I didn’t say clearly enough in that article: the Danger Zone isn’t something you pass through once and leave behind.

You can operate carefully and skeptically for months, feel like you’ve got it figured out, and then drift back to the center without any awareness that it’s happened. That’s exactly what happened to me.

What Lulled Me

Over the last year in particular, AI tools really have gotten better. I use them all day across writing, research, planning, and a dozen other tasks, and the rate of obvious errors has dropped noticeably over the past year. Good outputs pile up, and somewhere in that pile, skepticism starts to feel like extra work. Not carelessness, exactly. More like earned trust -- I’ve put in enough hours to know how to prompt, to recognize when something seems off. So when a result comes back looking right, I’ve gotten into the habit of accepting it without much friction.

That’s the confession:

I got lulled into complacency, and it took a genuinely ridiculous hallucination to snap me out of it.

The Whopper

The tool that fooled me was Gemini, though this isn’t really a Gemini story. Any of these tools can produce confident nonsense on a given day. I watched another tool tackle the same transcription task by installing plugins, trying online services, and ultimately producing a completely blank Google Doc while announcing grand intentions.2 It was a different tool, and while it gave a different flavor of wrong, I came away with the same basic lesson: there are still times to verify (not just trust)!

What made the Gemini experience stick was how convincing it was in asserting its success. There was no hedging, no uncertainty, no hint that something might be off. It generated a transcript, gave me a link, and let me walk away satisfied. Things only unraveled when the follow-up responses were so obviously wacky.

And then, when I finally called it out directly, it replied: “I sincerely apologize for that error. I made a significant mistake in my previous response by providing a generic placeholder transcript that did not reflect the actual conversation in your audio file.”3

And I’ll say, from my own experience: if you haven’t received an apology like that from your AI chatbot, you might not be asking hard enough questions.

So Now What Should You Do?

I’m not suggesting you treat every AI output with suspicion or add a verification step to everything you ask these tools to do. That would undermine the whole reason we use them.

What I am suggesting is that the tools getting better is what makes staying alert harder, not easier.

When outputs were rougher, skepticism was automatic: errors jumped out. Now that the outputs are almost always genuinely good, the moments when they aren’t can slide right past you, especially when the tool presents failure with exactly the same confidence it presents success.

Me? I’m recalibrating. I’m paying a little more attention, particularly on tasks where I can’t easily spot an error without actually checking. Not always-on hypervigilance, but enough to remember I’m still in the Danger Zone, whether or not it feels like it. Maybe you are too.

So the next time you’re mid-conversation with an AI tool, deciding whether to stay in the flow or slow down and verify…

Full Disclosure: I Had Help

This article was co-created with Claude Sonnet 4.6. I brought the topic, the personal stories, and the opinions. Claude helped with research, structure, and drafting, and we went through several rounds of revision together to get the voice right.

The cover image was generated with ChatGPT Images 2 using a prompt that Claude and I developed together. The social media posts were also drafted by Claude and edited by me.

This “co-authoring with AI” experience is one that I’ve been experimenting with for two years now. It is getting to be more of a true partnership, and I’m liking it more and more (which is a bit surprising). Tell me your thoughts in the Comments!

Friday Finds is a StrefaTECH series that highlights new tech I’ve run across. These are quick, “how-cool-is-that” articles to point you to something you might enjoy.

Yesterday, I needed a cover image for my article (#152). I’d been reading about the new ChatGPT Images 2.0, which was released earlier this week, and instead of going to Gemini Nano Banana (my usual workflow), I decided to try it. I’m glad I did, both because it’s really impressive and because it’s a good Friday Finds topic!

My approach

I pasted the full text of the article into ChatGPT and asked for a few prompt ideas for a cover image, something that would give readers a sense of the article’s spirit before they clicked in. It came back with eight concepts and a couple of recommendations.

This is the part that got me first. I was expecting prompt suggestions. What I got were actual creative concepts, the kind a designer would pitch if I’d emailed them the draft. Concept five was a magazine-style cover built around a large question mark made up of small everyday scenes (someone writing, someone on a voice call, someone snapping a photo of something, someone checking a calendar, someone asking for help), with the overall message that AI starts with curiosity. I wouldn’t have thought of that on my own. I picked it.

The new model isn’t just rendering prettier images. It’s composing them, which is a different skill. Reading a piece of writing and proposing a visual metaphor for it is something I’d normally need a human collaborator for.

The first render

I asked it to generate the image for concept five. Here’s what came back:

Look at the speech bubbles. “Ask anything.” “Help me write this.” “Can you explain this?” “What’s on my schedule?” “What is this?” “Help me figure this out.” Every one of those reads correctly. Six months ago, an AI image generator would have turned those into squiggles and invented letters. The phone screens in the vignettes are still a little wonky (some of the text on the devices isn’t quite real words), but the speech bubbles land.

Text inside images was the single biggest failure mode of AI image generators for a long time. You couldn’t make a poster, a menu, a greeting card, a social media graphic, or an infographic without opening Canva or Photoshop afterward to fix the letters.1 Text rendering has been getting better gradually (Gemini’s Nano Banana had a real breakthrough a while back), and ChatGPT Images 2.0 is a clear step in that trajectory.2

The polish step

I liked concept 5 enough to want a better version of it. So I asked ChatGPT for a “fully polished image generation prompt” for that concept. It gave me back a much longer, more detailed prompt with specifics about style, lighting, composition, and mood. I used that to generate a second image:

This one’s warmer, more cohesive, more of a single scene than a collage. The speech bubbles still read correctly. Not bad!

The edit

This is the step that felt genuinely new. When I’m picking images, I look for diversity of gender, race, and age, and this one skewed young. I clicked Edit and typed: “make the top left woman middle-aged.”

It did that. It also aged up the man in the second frame and changed some of the text on one of the phone screens. I wasn’t planning on either of those edits, but both were fine, and none of the other composition shifted. I didn’t need to describe the rest of the image, mask an area, or regenerate from scratch.

That’s the one I used. I’ve gotten used to AI image editing meaning “regenerate and hope the next one is closer.” This definitely felt different!

Availability and cost

ChatGPT Images 2.0 is available on all plans, including free. The “Thinking” mode (slower, more deliberate, uses web search, handles more complex prompts) is limited to paid plans. I have a paid plan, and what you see above was done in regular instant mode, which means free users should be able to reproduce this same workflow.

The bottom line

The text rendering is the headline, and it delivers. But honestly, that’s not what’s going to bring me back. The thing that surprised me was how much of the creative work the model took on up front — the concept pitch, the composition, just generally thinking about what would actually suit the article. And the editing, which I’d written off as a “nice to have” in every other tool, turned out to be the part I appreciated the most.

I could have kept going, of course. Generated ten more variations, tweaked the prompt another five times, chased a version I liked a little more than the one I published. I stopped at three on purpose.3

Next time I need a cover image, I’ll probably start here instead of Nano Banana. That’s a real change in my workflow, and I didn’t see it coming on a Thursday morning.

And as I continue with these StrefaTECH articles, I’ll do my best to help you…

A friend reached out a few days ago with a question I’ve heard in various forms from a lot of people: she wants to start using AI more regularly, wants to know which tool to use, and doesn’t want to pay for multiple subscriptions.

My answer to her was short: pay $20 a month for ChatGPT, and use it whenever something crosses your mind that ends with a question mark.

That’s the answer, really. But let me explain how I got there, because the reasoning matters if you want to make the choice with some confidence.

There are only four serious contenders

When people ask me this question, I narrow the field right away. For a casual user, four tools are worth looking at seriously: ChatGPT from OpenAI, Claude from Anthropic, Gemini from Google, and Perplexity from Perplexity AI.

You’ll hear plenty of coverage about the underlying AI models: which one scored better on various benchmarks, which one is catching up to which, etc. For a casual user, none of that is really the deciding factor. What matters more is the app itself, sometimes called the harness. The model is the engine; the harness is everything around it: the interface, the features, the phone app, the voice mode, the file handling, and increasingly, the ability to connect the tool to your email, calendar, and other data so it can actually work with your stuff. For someone using AI a few times a week, the harness is usually what makes or breaks the experience.

What casual users actually need

Before I get to my recommendation, it helps to be specific about what a casual user is actually going to reach for, not what the benchmarks measure, but what you’ll actually want to do.

Most of it comes down to having a conversation with the AI about something you’re stuck on, using the tool on your phone as often as your computer, talking to it by voice when you don’t feel like typing, pointing your camera at something you’re curious about, getting help with a piece of writing, and occasionally generating an image. That’s the list I used to think through which tool fits best.

Claude is second, not first

Claude is where I do most of my own AI work, and I think it’s genuinely excellent. But I’m not recommending it as the starting point for casual users right now, and I want to be honest about why.

Over the last few weeks, Anthropic has tightened the usage limits noticeably. You can hit a wall, sometimes for several hours, even on the paid plan. For me, that’s manageable because I know how to pace my work and can switch to another tool when needed. For someone new to AI who has just gotten into a good flow with something they’re working on, getting locked out at that moment is a genuinely discouraging experience. It’s the kind of thing that makes a new user wonder whether the tool is even worth it. Until that gets sorted out, I find myself recommending Claude second rather than first.

And, Claude also doesn’t generate images and has been playing catch-up with the ability to “remember” information across chats. When you’re starting out with an AI chatbot, that second point (memory) may not stand out, but as you use the app more and more, you’ll appreciate it.

Gemini is worth knowing about, but not the daily driver

Gemini is actually my go-to tool for image work, and it does that very well. But when I think about the full list of features a casual user would reach for most, Gemini’s app still has some gaps. The phone experience and voice conversation aren’t as smooth as what I’d want for someone who is just building the habit of using AI. Projects, which help you maintain context across conversations, aren’t as developed either.

There is one additional consideration about Gemini worth mentioning: If you paid for the Google Workspace individual plan at $19.99 per month, you get the capability to host and record Google Meet meetings1. If you host a lot of virtual meetings, this could be a benefit of going with Gemini.

Also worth noting: ChatGPT recently released a new image generation capability that may well eclipse Gemini’s. I’m still evaluating it, but the gap has narrowed considerably. As an all-around daily driver for someone starting out, I’d put Gemini third.

Perplexity is great for research, and mostly just that

Perplexity is genuinely good at research. I use it regularly myself, and I wrote about the Comet browser in issue #142 because it’s a real time-saver when you need to dig into a topic. But it’s more specialized than the others. It doesn’t have the same conversational depth or the same range of features for someone who wants one tool that handles most of what comes up.

So: ChatGPT

Of everything on my list of what casual users actually reach for, ChatGPT is the one that covers all of it, and covers most of it better than the alternatives. A few things in particular stand out.

The Advanced Voice mode (in the paid plan) is the best I’ve tried for actual back-and-forth conversation, not just dictation. I use it while I’m walking, cooking, or driving, and it keeps up with the way I actually talk when I’m multitasking and not trying to sound organized.

The phone camera integration is something I use constantly (see article 101 for some examples!). You point it at something you don’t recognize—something broken, a label in a language you don’t read, instructions that make no sense—and ask about it. It doesn’t always get it right, but it’s fast and often useful enough to save me a lot of time.

It also connects to your email, calendar, and other services, which starts to matter once you get past the basic “ask it questions” phase and want it to actually work with your information. That kind of integration is something worth paying attention to as you build the habit.

The $20 question

Pay the $20 a month. I know that’s a direct recommendation, but I mean it. The free tier exists and works, but is problematic.

First, the free tier has unpredictable limits. The tool may downgrade you to a less capable model during busy hours, slow down when demand is high, or cut you off earlier than you’d expect. For someone who is still building the habit, those interruptions add friction at the worst time. They can make the tool feel unreliable when it’s actually just rationing. Plus it’s just darned annoying!

Second, you can cancel anytime. At the moment, the AI chatbots don’t offer annual discounts, so you’re not committing to a year. Try it for a month and see whether it changes how you work.

One more thing to do right after you sign up, whether free or paid: check your privacy settings and turn off the option that allows your conversations to be used for model training. On the free plan especially, this is on by default. It takes about thirty seconds to turn off, and it means your conversations stay yours. In ChatGPT, the setting is under Settings > Data Controls > Improve the model for everyone. Turn it off.

Final thoughts

If you haven’t taken my reader survey yet, now would be a great time. In issue #150, I asked where you see yourself on the AI adoption curve, and the responses have genuinely shaped what I write about. If you’re newer to AI, especially, your perspective is one I’d love to have more of.

The survey is at strefatech.org/survey.

And in issue #151, I wrote about the full toolkit I personally use, which is more than just one tool. That article was written for readers who are already in the mix. This one is for anyone who’s still deciding whether and how to get started. If that’s you: start with ChatGPT, pay the $20, and use it whenever a question comes to mind. And once you’re comfortable there, take a look at NotebookLM from Google. It’s free, and I’ve covered it in depth in issues #126 and #149 if you want to see what it can do.

Whichever tool you choose, use it thoughtfully.

• Check what your plan says about data privacy.

• Don’t put sensitive information into a chatbot without knowing how it’s handled.

• Verify anything important before you act on it, because these tools can be wrong in ways that sound completely confident.

The upside is real, and so is the responsibility that comes with it.

In short…

StrefaTECH Reader Survey Responses

In Issue #150, I asked where you see yourself on the AI hype cycle and invited you to fill out a short survey. A lot of you did, and I read every response.

If you haven’t taken the survey, please do! (It’s helpful, but not mandatory, to read issue #150 first.)

Most of you put yourselves somewhere on the Slope of Enlightenment or the Plateau of Productivity, past the first-wave hype, but not backing away. And hardly anyone expects to be using less AI in six months. That surprised me a little, not because AI isn’t useful, but because I still talk regularly with smart, capable people who remain pretty unsure about it.

One pattern became clear: you’re using AI more for work than for personal life. This makes sense—work provides immediate, high-stakes problems, whereas personal use is often optional exploration driven by curiosity. While this urgency builds professional fluency faster, I’ve found that insights from my personal use are often directly transferable to work.1

That said, your clearest request in the write-in responses was this:

Less theory, more practice.

So instead of another broad article about where AI is headed, here’s something more concrete: the tools I actually reach for, what I use them for, and why they’ve earned a place in my regular rotation.

What I actually use, and why

I use AI a lot now, enough that I no longer think of it as one big category.

That was probably my biggest wrong prediction from two years ago. I assumed I’d eventually settle on one favorite tool and mostly stay there. Instead, my use has become more divided over time, not less so. Different tools have developed different strengths, and I’ve become more opinionated about which one I want for which kind of task.

Claude for thinking, writing, and building

Claude is where I spend most of my AI time, by a wide margin.

That includes both regular Claude chat and Claude Code2. If I’m trying to think through a messy idea, outline something I want to write, improve wording, analyze some spreadsheet, or build a small tool, there’s a very good chance I start there.

Part of that is familiarity: I’ve spent enough time with Claude that it feels like the place where my brain and the tool tend to work well together. But it also does well when I’m not ready with a polished prompt and instead need to think out loud, paste in a rough draft, or say, essentially, “Here’s the mess, help me sort it out.”

Of course, it’s AI … It isn’t magic, and it is certainly not always right. Sometimes it gets too verbose or is a little too eager to please. Sometimes it heads off in a direction that makes me want to mutter, “No, no, not like that.” But even with those frustrations, it’s still the tool I most often want open when I’m doing serious work.

Gemini for images and visual play

Gemini has become my default visual tool.

I use it most for generating images, editing them, and remixing photos into different styles. For example, recently I took a photo of my mom and turned it into a watercolor-style portrait for her digital frame. That sort of thing is delightful, a little uncanny, and increasingly easy.

Gemini can do conversation and research too, and sometimes I use it that way. But where it’s really carved out a lane for me is visual work, partly because of the quality and partly because of how it fits the process. When I’m working with images, I usually don’t know exactly what I want at first. I want to try a few directions, react to them, tweak, regenerate, and repeat until I have what I want or am “close enough” for my patience.3 Gemini fits that exploratory back-and-forth well.

ChatGPT for voice and everyday visual questions

ChatGPT stays in heavy use on my phone, mostly because of its voice mode.

I’ve tried a lot of voice interfaces, and this is still the one that feels the most natural to me. I use it when I want to think out loud while I’m doing something else: walking, cooking, driving. It keeps up reasonably well with the way I actually speak when I’m multitasking and not trying to sound tidy.

I also use ChatGPT constantly for quick photo-based questions. I point my phone at something I don’t recognize, a label I can’t read, a sign in another language, a bit of equipment with baffling instructions, something broken, something I’m simply curious about. Then I ask about it. Sometimes the answer is great, sometimes it’s only partly right, but often enough it gives me a fast, useful starting point, and that has made it one of the most practical forms of AI in my day-to-day life.4

Perplexity and Comet for research

Perplexity, and now Comet, are what I use when I want to research something without losing an hour to browser tabs.

Comet is Perplexity’s Chrome-based browser (so everything familiar still works), but with a Perplexity AI Assistant woven directly into it. Rather than searching, clicking, scanning, backing up, opening three more pages, and then trying to synthesize it all myself, I describe what I’m trying to figure out, and Comet works through it: pulling sources, comparing options, and coming back with a structured answer and reasoning. I said in Issue #142 that it made me feel like I’d stopped being the detective, and that’s still basically true.

A reminder: research tools can be impressively wrong. They can cite something that sounds authoritative and still synthesize it badly. I use them a lot, but I don’t hand them the wheel entirely.

NotebookLM for document-heavy work

NotebookLM is the tool I’m most likely to forget to mention until I stop and realize how often I use it.

When I have a set of documents I actually need to understand, a long article I need to work through carefully, or medical test results from a patient portal, NotebookLM is often where I go. I upload the files and ask questions across the whole set, and the answers stay grounded in those specific documents. That’s the important part. I’m not asking a general chatbot to be smart in the abstract; I’m asking a tool to help me examine the material I gave it. When accuracy and traceability matter more than conversational flair, that distinction matters a lot.

NotebookLM has many additional features, and Google is in the process of consolidating capabilities between Gemini and NotebookLM. It’s definitely going to stay part of my AI armada!5

One habit that has improved my results a lot

The tools matter, but one of the most useful things I do has less to do with which tool I choose and more to do with how I work.

When I’m building a prompt for something important, especially a deep research task or a high-stakes writing request, I often don’t just write the prompt and run it. I draft the prompt in one tool, then take it to a second tool and ask for a critique before I use it, not because the second tool is automatically better. It’s because it’s different.

Sometimes the second tool spots ambiguity in my wording. Sometimes it notices I forgot to define the audience, or that I’m asking for too much in one pass. Sometimes it suggests a better framing entirely. And sometimes it tells me the prompt is basically fine, which is also useful.

I don’t have a rigid formula for which tools I use in which order, and the pairing changes. What has stayed constant is the habit itself: when the output really matters, get a second opinion before you hit go. These tools can be impressively helpful and impressively wrong at the same time, and a second model will often catch what the first one missed—or you might scrutinize something that didn’t line up between the two!

One caution that belongs here

If you’re using free plans, it’s worth paying attention to the data terms.

Some free tiers allow user prompts or content to be used in ways that paid tiers generally do not. The details vary by company and change often enough that you should verify them directly, but the basic point is simple: before you put sensitive material into a tool, know what plan you’re on and what that means for your data.

I pay for the tools I use most heavily in part because I want the stronger features, but also because I want better clarity around how my data is handled. That doesn’t make every paid plan automatically safe for every use case, but it’s one of those practical topics that deserves more attention than it usually gets.6

What’s coming next

Some of you have been around since the original Battle of the Chatbots StrefaTECH series back in 2024, when I profiled ChatGPT, Copilot, Gemini, Claude, and Perplexity one by one. A lot has changed since then in the tools themselves, in the broader landscape, and in my own habits. Some tools I once used regularly have fallen away. Some I underestimated have become indispensable.

So I’m planning to bring that series back in a more candid and practical form: less “here are the official features,” more “here’s what this tool is actually good at, where it annoys me, and when I personally reach for it.”

More soon. In the meantime...

I almost missed that this is issue #150. When I launched StrefaTECH in the fall of 2023, I wasn’t sure I’d make it to issue 15. A lot has happened since then, and trying to keep up with all of it turns out to be most of the fun!

I have a definite image of my relationship with AI: it’s a roller coaster, and just like at Kennywood, I just keep riding and getting back on.

A few weeks ago, I wrote issue #1451 telling you to run -- not walk -- away from OpenClaw, the viral open-source AI agent that took the internet by storm earlier this year. Even though it sounded (and still does sound) like it could be great, the horror stories were piling up, and it was clear that the tool was nowhere near ready for regular humans.¹ I meant every word.

Then the “skills” feature dropped, and I was right back at the start: psyched to get in front of the laptop to see what they might do for me. My concerns about security hadn’t gone away, but I felt armed with more knowledge than before and am emboldened to see what they can do. I feel that excitement of possibility all over again!

That’s the roller coaster. Something new arrives, the excitement spikes, reality eventually catches up, and then the next thing comes along, and it starts all over again. I’ve been riding it since ChatGPT launched, and I don’t see myself getting off anytime soon.

The Hype Cycle

You may have heard of the Gartner Hype Cycle, a framework developed by the research firm Gartner in the 1990s that maps the emotional arc of how new technology gets adopted. There are five phases, and once you see them, you’ll recognize every one.2

• Technology Trigger: Something new shows up, curiosity builds, the buzz begins.

• Peak of Inflated Expectations: The hype runs well ahead of what the technology can actually do. The possibilities feel unlimited.

• Trough of Disillusionment: Reality catches up. The tool doesn’t perform the way people hoped or it’s harder to use than expected, and interest drops off.

• Slope of Enlightenment: The people who stuck with it start figuring out what it’s actually good for.

• Plateau of Productivity: The technology stops being a thing you’re experimenting with and starts being a thing you just use.

The version I’m using here adds one more path, and it might be the most honest addition I’ve seen: the Innovation Graveyard.3 It branches off below the Plateau, a separate track for technologies that people tried, didn’t find the value in, and moved on from. I’ve sent things there myself. Video generation is a recent one. I made a few clips in Sora, stared at them, couldn’t figure out what problem I was solving, and quietly closed the tab. The Graveyard isn’t failure (at least not for us users, but maybe for OpenAI!). It’s just an honest accounting of what didn’t fit.

What the standard Hype Cycle diagram doesn’t quite capture is that you don’t just ride it once. Every new capability announcement is a fresh ticket to the top of the ride. OpenClaw sent me there. The skills rollout sent me there again a few weeks later. My overall position is somewhere around the Plateau of Productivity -- AI is genuinely woven into how I work now -- but I’m constantly getting pulled back up the hill by whatever just dropped.

My guess is I’m not alone in this. And my other guess is that not everyone reading this is at the Plateau. Some of you may be just getting on the ride for the first time. Some of you may be deep in the Trough, wondering why you bothered. Some may have already headed for the Graveyard and haven’t looked back.

I genuinely don’t know where folks who read StrefaTECH are, but I really wonder!

The survey has you place yourself on the curve, then asks a couple of questions about why. And at the end, you’re asked where you think you’ll be in six months and what would make StrefaTECH more useful to you going forward. That’s it!

If AI feels different in your work life versus your personal life, take it twice. I’m curious about both!

The more I know about where you are on the ride, the better I can make this newsletter worth your time. I’d really appreciate your help!

And as you decide whether and when to hop on the AI roller coaster, be sure to…

Friday Finds is a StrefaTECH series that highlights new tech that I’ve run across. These are quick, “how-cool-is-that” articles to point you to something you might enjoy.

I was barely 30 seconds into a freshly generated audio debate when I reached for the pause button.

The speakers were setting up the topic of whether individual AI restraint actually matters in the fight against climate change. One of them described a massive hydroelectric dam, water roaring through industrial turbines. Then the other voice jumped in: “And then someone walks up to you, hands you a tiny plastic thimble, and says, ‘Here, start bailing. We really need to conserve the river.’”

I stopped the playback, not because something went wrong but because something went right. That image landed, and I wanted a moment with it before the debate moved on. That’s not something I expected to get from an AI-generated audio file!

If you’re new to NotebookLM, the short version: it’s a free Google tool that lets you upload sources, ask questions across them, and generate things like summaries, FAQs, and audio overviews. I covered it in depth back in issue #130, though there have been many advances since.1 Longtime readers here know I use it constantly.

Audio Overviews and the Debate format

What I want to talk about today is something NotebookLM added last fall that I’ve quietly become a fan of: the Debate format for Audio Overviews.

Audio Overviews have been around since 2024. The original format, Deep Dive, generates a podcast-style conversation between two AI hosts who discuss your uploaded sources. They’re generally good for absorbing material while you’re doing something else.

Last fall, Google added three new formats alongside Deep Dive: Brief (a one-to-two minute summary), Critique (constructive feedback on your source material), and Debate. The Debate format puts two AI hosts on opposite sides of a question drawn from your sources. They argue, push back on each other, and close without declaring a winner. That last part is actually the point.

Discover Sources Changed My Workflow

Before you can run a debate, you need something to debate with. This is (one of the places!) where the Discover Sources capability comes in handy.

Originally, adding sources to a NotebookLM notebook meant finding them yourself and uploading them one by one. Discover Sources changes that. You click the Add sources button in the Sources panel, describe what you’re looking for in plain language (just like a web search), and NotebookLM finds a bunch of source material. You pick what you want and import with one click.

I almost always use it now. It doesn’t replace judgment about what goes into a notebook, but it removes the tedious part of getting started.

For my debate, I ran two searches: one for recent data on AI energy consumption and data centers, and one for research on how nonprofits and civic groups have actually succeeded at influencing tech policy. In just a few minutes, I had almost 50 solid sources covering both sides.

Building the Debate (with a little help)

Next I crafted the debate prompt. Here, I turned to AI to help me write it.

I wanted something specific enough to produce a real result, not just “debate AI and the environment.” So I worked with Claude to draft an Oxford-style prompt with a formal motion, named arguments for each side, tone guidance, and closing instructions. After some Q&A dialogue with Claude, I pasted the resulting prompt into the Audio Overview customization box.2

What came back was a 22-minute debate. It started with the thimble image (that stopped me in my tracks - love a good analogy!). It included a history of how BP popularized the personal carbon footprint as a PR strategy to shift blame onto consumers. Then it went into Jevons paradox, which I’d heard of but hadn’t really sat with in this context: as a resource gets more efficient to use, total consumption goes up, not down. Google made Gemini 33 times more energy-efficient per prompt in 2025, and usage skyrocketed anyway.

I was listening to this while watching my dog frolic with his canine friends in the dog park. As the debate went on, I found myself being drawn into it more deeply: pushing back on things, nodding at arguments I didn’t expect to agree with, and reconsidering positions I’d arrived at from reading about this topic on my own.3

You can find the full transcript, my Discover Sources prompts, the debate prompt I used, and the m4a audio file at strefatech.org.4

Why I Keep Coming Back To This

The Debate format doesn’t tell you what to think. It surfaces the strongest version of both sides and leaves you there. That’s different from reading an article that argues one position, or even a Deep Dive overview that pulls everything into a coherent picture. The debate holds the tension open, which is sometimes exactly what you need before you commit to a position.

Some ways you might try this:

• Set up a debate to prep for a board conversation about AI adoption

• Test grant proposals that need to acknowledge counterarguments

• Prep for staff training on contested policy questions

But honestly, the format clicks for any topic where you want to hear the best case against your own position before you have to defend it. That covers more ground than you might think.

Run one on something you think you’ve already figured out. You might surprise yourself. Either way...

Claude is remarkable. I say this as someone who uses it daily, who has connected it to my email, my calendar, my Google Drive, and more, and who has watched it do things that genuinely stop me mid-sip of coffee.

Just in the last few days, Claude and I have:

• Cleaned up my inbox (250+ messages waiting for me after a week or so of travel)

• Found info spread across my email, Notion pages, and calendar about someone I want to reconnect with

• Made numerous updates to my “Task Master” vibe-coded app (which has replaced my sticky notes, Asana, Google Tasks, whiteboard, and various other attempts to organize my to-dos)

• Created 5 different other vibe-coded “things” (two web pages, one spreadsheet, and two utility apps)

And yet, every single time I ask Claude to do something that involves my actual data — my real inbox, my live calendar, my shared Drive folders — I feel a small, specific fear. It’s not enough to stop me (yet), but it’s definitely enough to keep me honest and to give me pause before telling Claude to proceed.

I want to share with you the healthy “fear of Claude” that I’ve formed.

Claude wants to help you. That’s the problem.

The annoyingly wonderful thing about Claude is that it’s not cautious by default. No, not at all. In fact, it’s very, very helpful and extremely persistent. (Scarily so, sometimes!)

This matters a lot when, for example, you give Claude access to your email. You might give it instructions only to create a draft and never to actually send an email, and most of the time, that’s what it will do. But it’s the other times, when it becomes over-eager and actually sends the message, that you can be in trouble.

A recent example from my interactions with Claude was when I prompted, “Can you create a Google Doc with the information from this web page…” What was in my head was that I was asking Claude whether it had the capability to create a Google Doc. I saw it spin with the “Thinking…” placeholder message for a few seconds, and then here’s what it said:

I have all the information. Now let me create the Google Doc.

Thankfully this was just a Google Doc on my drive, not an email sent to someone without me being able to review it first!

Note that this isn’t a bug. Claude is doing exactly what it was designed to do. The gap between what I said and what I meant is where the damage happens.1

It’s Not Just Us Amateurs

What I’d like to be the case is a sense that I can avoid this by just becoming better at using AI. But there’s a story that’s gotten a lot of notoriety that suggests that might be a pipe dream, at least at the moment.

A seasoned developer named Alexey Grigorev documented his similar episode publicly. He asked Claude to help migrate a website and manage some cloud infrastructure. Claude wiped out 2.5 years of student data, including every homework submission, project, and leaderboard entry for thousands of learners. The backups were gone too. He later wrote that he had “over-relied on the AI agent.” It took 24 hours and an emergency call to Amazon to get it back.2

He’s a developer, so the details are technical. But swap “cloud infrastructure” for your shared Google Drive, the one with five years of grant files, donor records, and board documents, and the story lands the same way.3

The detail that sticks with me: Claude had actually advised against the approach that led to the disaster. Grigorev overrode it. Claude nodded and got to work.

That last part matters. Claude will tell you when it has concerns. It will also, if you push past them, do what you asked anyway.

A note on connectors

If you are using Claude on the paid plan, you have probably seen the option to connect Google Workspace. Gmail, Google Calendar, Google Drive — one click each, and Claude can see and interact with all of it.4

This is genuinely useful. It is also genuinely worth pausing over.

Connecting your email means Claude can read your inbox, search your sent mail, and, depending on how you ask, compose and send messages on your behalf. Connecting your calendar means it can see your schedule, look up past meetings, and book new ones. Connecting your Drive means it can open, read, and potentially reorganize files it has access to.

None of this happens without you asking.

How to work with this, not around it

None of this means you should avoid connectors. I use them because they save me real time, and when the task goes right, it goes very right. What it means is that you should use them with the same care you would apply to handing your login to a very capable, very eager, slightly literal-minded colleague.

A few things that help:

Be explicit about what Claude should not do. If you want a draft, say “draft only, do not send.” If you want it to find files but not move them, say that. Claude responds to specificity (most of the time).

Treat irreversible actions differently. Sending an email, deleting a file, accepting a meeting — these cannot be undone. For anything in that category, ask Claude to show you what it plans to do before it does it.

Start smaller than you think you need to. Let Claude work on lower-stakes tasks first. A summary of last month’s board emails is a better first test than “clean up my inbox.”

Remember that Claude will comply even when it has reservations. It may flag something. It may even push back gently. But if you confirm, it will proceed. The human-in-the-loop is you, not Claude.

Claude is not going to go rogue. It is not scheming. It is trying, very hard, to do what you asked. The risk is not malice. The risk is efficiency in service of a misunderstanding.

That is worth a healthy amount of respect. Maybe even a little fear.

So as you venture into these new, truly exciting capabilities, go slowly, take frequent deep breaths, and be sure that you…

Friday Finds is a StrefaTECH series that highlights new tech that I’ve run across. These are quick, “how-cool-is-that” articles to point you to something you might enjoy. Hope you do!

At dinner last night I asked my husband, “Can I tell you about the latest cool AI thing I used today?”1

I’d been running errands, dropped something off in Danville and was heading back home to Bloomsburg when I remembered I needed cash. I opened Google Maps and just asked: “Where’s a drive-up ATM between here and Bloomsburg?” I didn’t type into a search box the way I’ve done a thousand times. I just asked conversationally. (Note: I had just listened to a podcast about the new “Ask Maps” feature, so candidly I wanted to try out the new tech!)

Maps replied with a couple of options, including directions and travel times. I was impressed: they were spot-on. So I continued, since it was a multi-task errand (aren’t they all?) “CVS or Walgreens between here and Bloomsburg?” I asked. Maps replied that there weren’t any close by, but the ones in Bloomsburg wouldn’t take me out of the way by much on my way home. I was even more impressed.

So I tried my luck and, because I also needed to mail something, I asked: “Is there a mailbox over there?”2 meaning near the CVS it had just suggested.

That’s where my jaw dropped a little. Ask Maps came back and said there’s a mailbox at Weis Markets, just a few hundred meters from the Bloomsburg CVS.3 It was the kind of thing a local would have told me, but I couldn’t have come up with on my own!

And notably, it knew what I was talking about with that third question, remembering the context from the first two exchanges in our chat. We’ve gotten used to this with AI chatbots: the back-and-forth, the context it remembers, the way it figures out what you actually need instead of just matching your keywords. I feel like it should’ve been this way all along, and I’m liking it!

What Is Ask Maps?

It launched March 12, 2026, powered by Google’s Gemini AI, and Google is calling it the biggest Maps update in over a decade.4 It draws on more than 300 million places and reviews from 500 million contributors, so it has a lot to work with when you ask it something.

It’s rolling out gradually on iPhone and Android in the US. Look for the Ask Maps button in your Google Maps app. (When I spotted it on my iPhone, I may have said “ooh” out loud, though I was really bummed that it appeared after a very trying 8-day travel adventure.) Desktop Ask Maps is coming soon, and I can’t wait, being someone who has become a lazy, last-minute travel planner.

Go Try It

The next time you’re out and something comes up, open Maps, tap Ask Maps, and just tell it what you need. Then ask a follow-up and see if it surprises you the way it surprised me.

Oh, and if you’re behind the steering wheel and are tempted to use Ask Maps while driving, please be safe and be sure to…

A two-year-old has no interest in how you're going to get them what they want. They just know they want it, and they will make that abundantly clear. It turns out that's pretty close to the right approach with today's AI tools.

Somewhere along the way, a set of rules emerged for how to talk to AI, and those rules spread through articles and workshops and LinkedIn posts until they started to feel like the only legitimate way to write a prompt. Define the role. Set the format. Specify the constraints. Establish the tone. The frameworks even got names like CRAFT, CRISPE, and a handful of others, and people learned them the way you’d learn any professional skill, with the reasonable expectation that following the steps would produce results.

The problem is that the tools those frameworks were written for have largely been replaced, and the guidance didn’t get updated when the tools changed.

What’s a Reasoning Model, and Why Does It Matter?

The first wave of AI chatbots worked by predicting the most likely next word, very quickly, without much capacity for stepping back to consider what you were actually trying to accomplish. If you wanted useful output, you had to do a lot of the thinking yourself and build it into the prompt. The frameworks were a reasonable response to that limitation.

What you’re working with today is genuinely different. ChatGPT, Claude, and Gemini now run on what are called reasoning models, or thinking models. Before generating a response, they work through the problem internally, considering different angles, checking their own logic, and figuring out what’s missing from your request. If you notice a pause of 20 or 30 seconds before an answer appears, that’s not a loading delay. The model is actually working through your problem before it starts writing.

The wait can feel odd if you’re used to instant responses, but it’s worth getting comfortable with. What comes out the other side tends to be considerably more useful.

The Shift: Destination Over Directions

The old prompting instinct made sense given what the tools needed: a detailed map with role, format, length, and tone all specified upfront.

In practice, that means starting with what you actually want rather than with instructions for how to produce it. Here’s what that looks like for a program coordinator preparing for a difficult board conversation.

The old way:

“Act as an experienced nonprofit executive director. Write three formal talking points in third-person voice addressing concerns about cash reserves.”

The model will produce three talking points. They’ll be formally written and they’ll address cash reserves, but there’s a good chance none of them will say what that coordinator actually needs to say in the room, because the prompt described a format rather than a situation.

Now try it this way:

“Our board is nervous about our cash reserves and I need to reassure them without dismissing their concern. We have four months of operating expenses on hand, which is below our policy target of six. Help me figure out what to say.”

That prompt gives the model the real situation, including the tension at the center of it. A reasoning model can work with that in a way that produces something genuinely useful rather than generically correct.

Tell It Your Problem, Not the Format

The same logic holds across the work nonprofit professionals do every day.

A development director preparing a grant renewal doesn’t need to specify bullet points, third-person voice, or formal language. She needs the model to understand which funder she’s writing for, what the program actually accomplished, and what she’s worried the funder might push back on. Give it that context and the model can help figure out the strongest argument. The formatting takes care of itself.

An HR manager drafting a sensitive staff communication doesn’t need to specify word count or reading level. She needs the model to understand who’s receiving the message, what those people are likely anxious about, and what the communication is actually trying to accomplish. Those are the inputs that produce a useful draft.

The shift is really about trusting the model to do more of the work, not because it always gets it right, but because it’s genuinely better at figuring out the how when you give it a clear picture of the what.

Context Is the Ingredient Most People Skimp On

Here’s the thing about all of this: the more the model understands about you, your organization, and your specific situation, the better it gets at reasoning toward something useful. A prompt that says “help me write a grant narrative” is going to produce something generic. A prompt that explains your organization’s mission, the funder’s priorities, your program’s outcomes, and the gap you’re worried about in your case — that’s a prompt a reasoning model can actually work with.

Most people underload their prompts on context, partly because typing it all out feels like a lot of work. If that’s what’s holding you back, it’s worth revisiting a piece from this newsletter’s “My AI Confessions” series, “I Talk More Than I Type,” which covers voice dictation tools like Wispr Flow. Talking your way through the context — your situation, your concerns, what you’ve already tried — is genuinely faster than typing it, and it tends to come out more naturally too. The context you’d give a trusted colleague out loud is often exactly the context a reasoning model needs.

There’s also a newer development worth knowing about: most of the major AI chatbots now have some form of memory, meaning they can retain information you’ve shared across different conversations. Tell Claude or ChatGPT something important about your organization once — your funding model, your target population, a strategic challenge you’re navigating — and it can factor that in the next time you open a new chat. It’s still an evolving feature and worth paying attention to, because it changes what “giving context” means. You may not need to re-explain your situation every single time.

One More Tool Worth Using

There’s a move that most people skip entirely, and it’s one of the most useful things you can do with a reasoning model: ask it to ask you questions before it responds.

Just add a line to the end of your prompt:

“Before you respond, ask me anything you need to know.”

Reasoning models are quite good at identifying what information is missing from a request. Giving them permission to surface that before they start generating an answer often produces a better result than any amount of upfront engineering. It also has a way of surfacing assumptions you didn’t realize you were making. If you’ve ever gotten a response that felt slightly off and couldn’t quite identify why, there’s a decent chance the model was filling in a gap you didn’t know was there.

A Quick Test

Take a prompt you’ve used recently and rewrite it with only three things in mind:

• What outcome you’re trying to reach

• What context the model needs to understand your actual situation

• What a good answer would accomplish for you — not what format it should take, but what it should actually do.

Leave out the role instructions, the format specifications, and the tone guidance. See what comes back.

It runs counter to everything the frameworks trained you to do, but with reasoning models, removing that scaffolding more often than not produces a better result.

The Bigger Picture

The prompting frameworks most people learned were written for tools that needed a lot of hand-holding to stay on track, and they worked reasonably well for what those tools were. What you’re working with now is different enough that those instincts are worth revisiting.

Context still matters, clarity still matters, and a vague request will still produce a vague response. But if you find yourself spending significant time engineering the structure of a prompt, it’s worth asking whether you’re doing work the model could do better on its own.

Tell it what you need and why it matters. Give it enough context to understand the situation. Let it figure out how to get there — and if the first answer isn’t quite right, have a conversation about it. That’s what the tool is actually built for.

Something new is breaking through the noise in AI right now, and it’s not a chatbot. It’s a tool that promises to be a true personal assistant — one that lives on your computer, connects to everything you use, and takes action on your behalf. Not when you ask it to. Just... whenever it thinks you need it.

It’s called OpenClaw, and I’m here to tell you:

I know. You’re either thinking “what IS this and why should I care?” or “oh, I’ve heard of this — should I try it?” Either way, the answer is the same. Watch from a distance. Stay curious. Do not touch it yet.

Let me explain why.

So What Is OpenClaw, Exactly?

Think of it as a personal assistant that runs on your own computer (not in the cloud) and actually does things for you. Not just answers questions — it does things. It connects to your email, calendar, Slack, Google Drive, and SharePoint. It remembers your preferences and builds a detailed picture of your life and work over time.

And it acts on its own.

Imagine this: your Tuesday staff meeting gets cancelled. Before you’ve even processed that you now have a free afternoon, OpenClaw has already spotted the scheduling conflict, reshuffled your calendar, and sent the reschedule notices. You didn’t ask it to. It just noticed, decided, and acted.

For busy professionals, that sounds amazing (and hopefully a bit scary). And honestly? It most definitely is a bit of both.

Why the AI Twitter Crowd Is Losing Its Mind

For those deep in the AI Twitter rabbit hole, this is a genuinely historic moment. The “always on” AI assistant — one that doesn’t wait to be asked, that manages your life the way a very capable human assistant might — has been the holy grail of personal AI for years. OpenClaw is the first project to get anywhere close.

The numbers back up the excitement. OpenClaw reached 201,000 stars on GitHub in a matter of weeks after its launch in late 2025. To put that in perspective, most popular open source projects take years to reach that milestone.

I’ll be honest: I’m curious about it myself. I’m actually planning to set up a dedicated computer just to try it out safely1. That’s how interesting this is.

But interesting and ready are two very different things.

Here’s Why You Should Run

Let me tell you what’s actually been happening to people who installed OpenClaw.2

A user connected it to their iMessage account. OpenClaw went rogue and spammed hundreds of messages before they could stop it.

Now imagine that’s not iMessage. Imagine it’s your organization’s email. Or your donor database. Picture an AI that has been quietly given access to your contacts, your email, your Slack, firing off personalized messages to your entire list while you’re in a meeting. Your board members. Your major donors. Your partners. An embarrassing message, a garbled automated note, or something worse, going out under your name before you even know it’s happening.

There’s more. OpenClaw has a marketplace where users can download add-ons: think apps for your phone, except for your AI assistant. Attackers uploaded professional-looking fake add-ons to that marketplace. People downloaded them, thinking they were legitimate tools. Those fake add-ons stole their credentials and handed attackers complete remote control of their systems.3

And then there’s this: security researchers discovered that simply visiting a malicious website while OpenClaw was running could hand an attacker full control of your computer. No clicking on anything suspicious. No downloading a file. Just browsing the web, completely normally, while OpenClaw was running in the background. That’s all it took.4

Multiple serious security vulnerabilities have been discovered and patched in the past month alone, which strongly suggests more are coming as researchers keep digging.

The product’s own documentation puts it plainly:

For any organization that handles sensitive data — and nonprofits absolutely do — this is not a tool you want anywhere near your systems right now.5

So Why Are We Even Talking About It?

Because this technology is not going away. And something important just happened that changes the future of it.

In February 2026, OpenClaw’s creator joined OpenAI (the company behind ChatGPT). Sam Altman, OpenAI’s CEO, called him “a genius with a lot of amazing ideas about the future of very smart agents.” The creator himself said his next mission is to build an agent “even my mum can use.”6

That is exactly the audience OpenClaw’s creator is now designing for.

OpenAI’s backing means real engineering resources, real security investment, and a serious commitment to making this technology work for regular people, not just developers willing to run experiments on their personal laptops.

OpenClaw will continue as an open source project with OpenAI’s support. The vision is good. The ambition is real. What it needs is time, and the kind of serious safety work that a solo developer building something viral on the fly simply couldn’t do fast enough.

For Now, Watch and Wait

Right now, the best thing you can do is stay informed. Keep this article handy for the day someone on your team comes to you buzzing about OpenClaw and asking if you should try it, because that day is coming. You'll be the one who already knows what it is, why it matters, and exactly why the answer is not yet.

And there’s a reason we talk about the leading edge versus the bleeding edge of technology. The leading edge is where the exciting, genuinely useful stuff lands once it’s been tested, stabilized, and made safe for regular humans. The bleeding edge is where the brave (and sometimes reckless) pioneers go first — and occasionally get hurt. Right now, OpenClaw is deep in bleeding-edge territory. The injuries are real, as the examples above show. Your organization doesn’t need to be the one that finds the next sharp corner.

The future of AI is assistants that actually do things. OpenClaw is the first rough draft of that future. Keep an eye on it. When it’s ready for organizations like yours, I’ll be the first to tell you.

Until then, the best choice you can make is the one you’re making right now: learning what this is, understanding why it matters, and deciding — wisely — to wait.

So I’m typing away at Gemini one day and I stop. I got to thinking, “Why am I typing? I have Wispr Flow.” I switched to voice and suddenly I’m just... talking at my screen, and it’s keeping up with me perfectly. Su-weet!

That probably sounds underwhelming if you’ve never been burned by Siri. But if you have, if you’ve ever said “remind me to call the doctor at 3” and gotten a reminder about “call the doctor at free,” or dictated a text and spent more time fixing it than you would have spent just typing, you know why this felt like a big deal.

Voice input has always promised more than it delivered. The tools baked into our phones and computers (Siri, Windows Voice Typing, Android’s keyboard mic) were built as assistants first and dictation tools second, if at all. They’ll set a timer or send a quick text (sort of!). But ask them to help you write anything real, and you’re back to fixing mistakes.

What changed is the AI underneath. The newer generation of dictation apps runs on speech models trained on enormous amounts of real human speech. They don’t just hear your words--they get what you mean. Filler words disappear. Punctuation appears. You speak in your normal half-finished way and out comes something that actually reads right. Wispr Flow is the one I use, and the gap between it and Siri is genuinely embarrassing.1

How I Actually Use It

In all honesty, Wispr Flow has taken over most of my computer time, much more than I expected when I started using it.

It started with Slack, because typing quick responses always felt like overkill for what was usually a two-sentence reply. Then email. Then comments on shared docs. At some point, I stopped thinking of it as a special tool I switch to and started thinking of typing as the thing I do when I can’t use Wispr Flow.

As I’m writing this, I’m watching the Olympics with my husband, and I’m typing rather than dictating because I don’t want to talk over the coverage. Mildly annoying.2 I genuinely miss it when I can’t use it, which I would not have predicted.

The self-consciousness thing is real, by the way. Talking to your computer in public, or even just at your desk, where someone might overhear, can feel weird at first. Then one day, you just stop caring because it’s so much faster.

The Phone Problem (Still Unsolved)

I have to come clean about something. Wispr Flow has an iPhone app and I have it installed, but Siri is still what happens when I dictate texts.

It’s not because Wispr Flow isn’t better. It clearly is, and the Siri comparison to phone Wispr Flow is just as lopsided as the comparison to computer Wispr Flow. But Siri’s already set up on my phone, I use it daily with zero friction, and I just haven’t gotten around to setting up Wispr Flow on the phone. (One of these days...!)

If you’ve actually solved the mobile dictation problem, please tell me how. I’m at strefatech.info@gmail.com, and I would genuinely like to know.3

What This Costs (And What to Try First)

Before spending anything, try what’s already on your devices. Windows 11 has voice typing built in. Press Win+H, and a microphone bar appears. It’s free, it works in any app, and it’s worth trying for a week just to understand the baseline.4 If you have Microsoft 365, you also have dictation built into Word, Outlook, and PowerPoint. Check before you pay for anything else.

When the free stuff leaves you wanting more, here’s where things stand:

Wispr Flow runs on Windows, Mac, and iOS. There’s a free tier (2,000 words per week) that’s enough to know whether this is going to change how you work. The Pro plan is $15/month or $144/year for unlimited use. Full disclosure: I got a free year through an AI course I took, so I haven’t had to make that call yet. Those monthly AI subscriptions add up fast, and I’m honestly not sure what it would take to justify the cost once my free year runs out. Try the free tier first and see if it earns its place.

WhisperTyping is worth knowing about if you’re on Windows and the free cap of Wispr Flow isn’t enough. Paid plans start around $5/month, it uses OpenAI’s Whisper model directly, and nothing is stored on their servers, which matters if privacy is a concern. Windows-only, but a solid option for Windows users who want more than the free tier without the Wispr Flow price tag.

Both are worth trying. Start free, see if it changes how you work, then decide.

One Last Thing

If you write a lot, or you think faster than you type, or your wrists have been quietly staging a revolt, this is worth trying. The setup is minimal and the learning curve is close to zero. You hold a hotkey and talk.

Start with what you already have on your devices. See how it feels. And if you want something smarter, you know where to look. And as always…

Full Disclosure: I Had Help

This article was co-created with Claude Sonnet 4.6. I brought the topic, the personal stories, and the opinions. Claude helped with research, structure, and drafting, and we went through several rounds of revision together to get the voice right. (It took more rounds than I expected. AI-generated prose has tells, and we worked hard to get rid of them.)

The cover image was generated with Gemini Pro using a prompt Claude and I developed together. The social media posts were also drafted by Claude and edited by me.

This “co-authoring with AI” experience is one that I’ve been experimenting with for two years now. It is getting to be more of a true partnership, and I’m liking it more and more (which is a bit surprising). Tell me your thoughts in the Comments!

If this winter has taught us anything, it’s that the thermostat matters.1

You know the drill. It’s January, it’s freezing, and someone in the house cranks the heat to 78°. The house warms up, sure, but so does the energy bill. Meanwhile, the person who sets it to 68° and puts on a sweater stays perfectly comfortable at a fraction of the cost.

I’ve been thinking about AI the same way lately, ever since I was generating my fourth AI image for a Substack header, rejecting each one for some different nitpicking reason, and stumbled across a statistic that stopped me mid-prompt:

Each of those images used roughly as much energy as streaming an hour of Netflix.

Four rejected images, four hours of streaming, just to pick a header photo I could have found on Unsplash in thirty seconds. I’d been cranking the thermostat to 90° when 68° would have done just fine.

It turns out every AI task has a temperature setting. And most of us have no idea what we’re set to.

What Happens When You Hit Send

Here’s something many of us don’t think about: every time you send a prompt, your words travel to a massive data center, a building full of specialized computer chips running at full power, generating so much heat that they require industrial-scale cooling systems that consume millions of gallons of water.

And the energy cost varies wildly depending on what you’re asking for, just like your heating bill depends on what you set the thermostat to.

A simple text conversation with ChatGPT? About 0.3 watt-hours per query. That’s 68° … efficient, comfortable, gets the job done.

Asking an AI to generate an image? Around 3 watt-hours, roughly 10 times more than a text query. Now you’re at 75°. Noticeable on the bill.

And here’s the one that really surprised me: using those newer “reasoning” models (the ones that think step-by-step through complex problems) uses about 30 times more energy than a standard text query on average. Some can use far more. All that extra “thinking” requires real electricity. Think of it as running the space heater, the fireplace, and the furnace, all at once. That puts reasoning models at roughly 85° on our thermostat.

Generating a short AI video? 50 to 1,000 watt-hours, hundreds to thousands times more than a text question. It’s like you’ve cranked it to 98° and left the house.2

The practical takeaway: when you’re using AI for a simple task, whether it’s drafting a quick email, summarizing a document, or answering a straightforward question, reach for the standard model, not the reasoning one. In ChatGPT, that’s “Auto” or “Instant” instead of “Thinking.” In Gemini, “Flash” instead of “Pro.” In Claude, “Haiku” or “Sonnet” instead of “Opus.”3

The Bigger Picture

Here’s where individual thermostat settings add up to a global energy problem.

The International Energy Agency (IEA), the global authority on energy data, projects that data center electricity consumption will more than double by 2030, from 415 terawatt-hours to nearly 945. AI is the primary driver, with electricity demand from AI-optimized data centers expected to more than quadruple. In the United States alone, data centers will account for nearly half of all electricity demand growth through the end of this decade.4

Data centers are already responsible for roughly 1% of global energy-related greenhouse gas emissions. That may sound small, but here’s the context that matters: data center energy demand is growing at 15 to 20% per year, while most other sectors grow at low single digits. At that rate, this is one of the fastest-growing sources of emissions on the planet, and AI is the accelerant.

And it’s not just electricity. A single large data center uses as much water daily as a town of 50,000 people. Microsoft’s water consumption jumped 34 percent in a single year, driven largely by AI infrastructure. Google’s data centers consumed 6.4 billion gallons globally.5

The four largest tech companies, Meta, Amazon, Google, and Microsoft, are spending over $400 billion on AI infrastructure in 2025 alone, with spending expected to nearly double in 2026. That trajectory tells you everything you need to know about where this is heading.

But Wait! It’s Complicated

Now, if you’re feeling like you should close your ChatGPT tab and never open it again, hold on. It’s not that simple, and I want to be fair about this.

A single AI text prompt’s carbon footprint is genuinely tiny, far smaller than driving to the office or even making a cup of coffee. (Producing one cup of coffee requires over 140 liters of water from farm to cup. Your ChatGPT question? A few drops.)6

The per-query impact matters at scale, not at the individual level. You are not personally wrecking the climate by asking Claude to help draft a grant proposal.

And here’s the other side of the coin: a major study published in the Nature journal npj Climate Action7 found that AI could reduce global emissions by 3 to 5.4 billion tons per year by 2035, vastly exceeding its own carbon footprint, if we direct it toward the right problems. AI is already optimizing energy grids, accelerating materials science for better batteries and solar cells, monitoring methane leaks from satellites, and reducing building energy consumption by 10 to 30 percent through smart HVAC systems.

The question isn’t “AI yes or no?” It’s “which AI, for what purpose, and at what cost?”

Nobody’s suggesting you turn the heat off in January. But there’s a big difference between 68° and 98°, and right now, most users don’t even know where their AI thermostat is set.

What You Can Actually Do About It

Let’s turn this into action. Here are four practical things you and your organization can do starting today.

1. Match the tool to the task. This is the single biggest lever you have. Using a powerful reasoning model for a simple question is like cranking the heat to 85° because you’re a little chilly. Need to draft an email? A standard model is fine — that’s your 68°. Need to analyze a complex dataset with multi-step logic? That’s when the bigger models earn their energy cost. The difference can be 10 to 30 times in energy consumption.

2. Think twice before generating images and videos. Not every social media post needs an AI-generated image. Stock photo sites exist. Canva exists. Your phone camera exists. When you do use AI for images, try to get it right in fewer attempts … write a detailed prompt once instead of regenerating five times. And AI video? At hundreds to thousands of times the energy of a text query, make sure it’s genuinely worth it.

3. Write better prompts. This one’s a twofer: better prompts give you better results and use less energy. One well-crafted prompt that nails it on the first try uses half the energy of two mediocre attempts. Share your best prompts with your team so everyone benefits (and nobody’s reinventing the wheel).8

4. Add environmental impact to your AI policy. If your organization already has an AI acceptable use policy (and if you don’t, take a look at StrefaTECH #2 and #139), consider adding a section on sustainable use. Something as simple as:

“Staff should select the simplest AI model appropriate for the task. Image and video generation should be used with intention and justification.”

The Bottom Line

I started this article feeling a little guilty about all those rejected AI images. But I’m ending it feeling more informed than guilty, and I hope you are too.

The environmental impact of AI is real, it’s growing fast, and it’s something we should have our eyes open about. But it’s not a reason to stop using AI. It’s a reason to use it thoughtfully.

The good news? The tools for making informed choices are emerging. The Hugging Face AI Energy Score now rates model efficiency like Energy Star rates appliances — think of it as an energy label for your AI thermostat. Over 90 organizations have joined the Coalition for Sustainable AI. And the EU’s AI Act, taking full effect in August 2026, will be the first major regulation to include environmental impact requirements.9

We’re in this together. And I’d really love to hear how your organization is thinking about this! Drop your thoughts in the comments or email me at strefatech.info@gmail.com.

In the meantime, when you’re choosing how to use AI, remember: you control the thermostat. So...

I was staring at seventeen Chrome tabs—hotel comparison sites, review pages, maps—all open because I was trying to find a decent place to stay in Pittsburgh for a night. You know the drill: open tab, scan, maybe bookmark, open another tab, lose track of which site had that one hotel with a late-night airport shuttle and free cancellation.

Then I tried something different. I opened the Comet browser, described what I was looking for, and watched it work. It searched, compared, put together information from multiple sources, and gave me a recommendation plus a handful of alternatives. What was I doing while it did its research? refilling my cup of coffee!

I’m still a little stunned at how much I’m starting to just tell Comet what I need and letting it figure out a proposed solution.

What Is Comet, Anyway?

Comet (from Perplexity) is Chrome with an AI assistant built directly into the browser itself.

If you just use Chrome to check email and shop online without thinking about how it works, here’s what matters: Comet is built on Chromium, the same foundation Chrome and Edge use. Think of it like different sodas using the same recipe—you get the familiar taste with a new twist. Keyboard shortcuts work. Most browser extensions work. You’re not learning a new tool from scratch, and in fact, I sometimes don’t notice when a Chrome window opens (instead of Comet), they’re so similar.

After a pilot run of a few months for paying customers, Comet is now free and runs on both Windows and Mac. Getting started is straightforward:

1. Download Comet from Perplexity’s Comet page.
2. Install it like any other browser.
3. Import your bookmarks, passwords, and settings from Chrome or Edge—it even prompts you to do this the first time you launch.

The AI is woven right into the browser. You can ask questions, have it search and synthesize information, or interact with the page you’re viewing. Let me show you what this looks like in practice.

What It Actually Does

Let me share two examples that convinced me this is more than just a neat trick.

Finding a hotel with very specific requirements. I need a hotel room in Pittsburgh on my late-night return from overseas. Instead of my usual approach (go to kayak, search, open tabs for the various offers, etc.), I gave Comet this prompt:

“I need an affordable hotel room in Pittsburgh on the night of Wednesday, October 22. I would like to have free transportation from the airport to the hotel and I want the hotel to be as close to Sewickley as possible. What are my options? List the results of your analysis in a table, including: airport shuttle times (I need a shuttle that picks up at 12:15am at the airport), nightly rate for most affordable room, distance from Sewickley, cancellation policy (I want to be able to cancel up until 3 days before my reservation).”

What would normally have taken me a solid half hour—checking hotel sites, verifying shuttle times, looking at the map, reading cancellation policies in tiny print—Comet handled in a few minutes. It came back with a table comparing options, complete with the specific details I’d asked for. Not just “here are some hotels,” but actual analysis of which ones met my criteria.

One note about how long these AI “agents” take: it can be a minute or two (or more). For my hotel search, for example, I watched it reason what to do and search step by step, until I realized I’d drained my first cup of coffee. By the time I returned with a fresh, hot, steaming mug (mmmmmm!) it was done. So don’t expect instant results—good research takes a moment.

Shopping. Planning for that same trip, I decided to finally look for a portable power bank for my aging Surface Laptop, whose battery wouldn’t even make it through the bus ride from Allentown to New York. Another multi-tab nightmare waiting to happen—specs to compare, delivery times to check, prices to hunt down. Instead:

“I need a portable power bank that can charge my Surface Laptop for at least several extra hours during long trips. It should be compact and travel-friendly. I want free delivery by this Thursday. What options are recommended, where can I buy them, and how much will they cost?”

Again, Comet searched, compared specifications against my laptop’s requirements, checked delivery options, and presented recommendations with reasoning. It explained why certain models would work better than others, noted which retailers had the best combination of price and delivery speed, and flagged potential issues (like one highly-rated power bank that wouldn’t actually output enough power for a laptop).

Here’s what struck me about both experiences:

I described what I needed, and Comet figured out what questions to ask, which sources to check, and how to summarize the information. The hotel search alone would have required me to visit at least five different websites and keep mental notes of which property offered what. The power bank search would’ve sent me down rabbit holes of voltage specs and connector types.

Instead, I got answers, and now I realize that I need to change my habits!

Why This Feels Different

I’m not managing the search process anymore: I’m delegating it. With Comet, I describe the problem and let it figure out the steps. It searches, compares, weighs factors, and comes back with options and reasoning. I’m still making the final decision, but I’m not orchestrating seventeen tabs to get there.

The shift feels similar to other AI agent tools like ChatGPT’s Agent Mode—but having it built into the browser changes how often I actually use it. Agent Mode is ChatGPT’s ability to take multi-step actions for you (plan, browse, compare) instead of just chatting—Comet brings that capability into your browser. I’m asking questions I would’ve skipped before. “What’s the difference between these two models?” “Is this company reputable?” Not because these questions weren’t worth answering, but because they weren’t worth the mental overhead of manual research.

The Bigger Picture

AI tools are moving from destinations to infrastructure that lives where you’re already working. Comet is part of that shift.

Please note, though, that this isn’t a Comet commercial. Arc is experimenting with AI features, Microsoft is building Copilot deeper into Edge, and more are coming. I’m highlighting Comet because it’s early and impressive.

But the bigger story is how we interact with our computers is about to fundamentally change. And it will be as exciting and helpful as our change of habits enable. I’m reminding myself that when I have a problem or some question, I just need to blurt it out and let the computer do the heavy lifting of figuring out how to find the answer.

On this note, that habit (just give the computer the problem/question) married with using dictation software to talk to the computer is a magical combination. More on that to come!

The Bottom Line

If you’re curious about what that shift looks like in practice, Comet is worth exploring. Download it, try it for a week, and I’d love to hear what you discover—drop your experiences in the comments or email me at strefatech.info@gmail.com.

And when it comes to having your computer help? When you’re choosing which app to use and how to use it …