You can add music files, play and skip tracks, see album art, even get basic playback stats.

I could use it as my only player, if I wanted to.

I can’t code. But I did build this, with ChatGPT leading the way.

Honestly, I don’t know much more about coding than I did before the project started. But I did learn to think through a user experience—where things go and why, and what to sacrifice to keep things simple.

Prepare for things to change midstream

Because the AI space moves so fast, what worked to build software yesterday might not apply today. For example, I built a ChatGPT project with all the governing documents to guide it. ChatGPT wrote code, piece by piece, that I would paste into Apple’s iOS app building software, Xcode.

That worked ok, but was pretty slow.1

Then, partway through the build, ChatGPT’s Codex suddenly appeared inside Apple’s Xcode application. Apple just slapped it right in there. Now you didn’t have to bring code over from ChatGPT. You could just … do stuff … right inside Xcode.

You have to be flexible, because the AI is advancing so fast. It’s like you take off for a trip in a prop plane, and halfway through the flight jet engines just sprout and start firing.

“One-shotting” is mostly hype

You hear a lot about apps being built in a single pass—”one-shotted”—where the AI takes an initial prompt and builds the software in one swoop.

There was no one-shotting my app, that’s for sure.

I spent hour after hour making small updates, running back to ChatGPT with new errors and issues, and correcting code before moving forward. Sometimes I had to take several2 different runs at one issue to resolve it.

We are not in the microwave era of codeless development. Not even the air fryer era.

Vibecoding is in its convection oven era. You can cook up pretty cool stuff, but it’s going to take awhile.

The biggest lesson: focus on small annoyances

A music app was too big a swing. Yes, it works. No, it’s not as good as many existing solutions. It solves no unique problems3.

For most of us, “vibecoding” should revolve around small, specific issues or use cases without a solution.

For example, what pisses me off about modern music is what Spotify is taking from people and artists: intentional listening.

The smaller apps I’m working on now fight that head on. They’ll be tools for contemplation, for intentional listening, and intentional discovery.

A smaller slice, taking on a major annoyance. That’s the key.

Subscribe now

The streaming media utopia they sold us—own nothing, access everything—isn't holding up.

Video streaming costs rose nearly 20% last year. While we once paid for unlimited, commercial-free access to movies and TV shows, well, now all the services have ads to accompany our price increases.

And music?

Spotify is being overrun with AI songs, by design, and AI is furiously shoveling them at you to improve profit margins. Albums and artists appear and disappear off the services.

Streaming listeners are losing their relationships with artists:

In the lean-back listening environment that streaming had helped champion, listeners often weren’t even aware of what song or artist they were hearing. As a result, the thinking seemed to be: Why pay full-price royalties if users were only half listening?

Sliding music into the background is more profitable than serving an engaged audience.

The subscription media model has one fatal flaw: they can change the deal whenever they want. And they do.

Against that backdrop, vinyl records have made a comeback. CD1 sales are also up, led surprisingly by Gen Z, who see CDs as both vintage and practical.

I’ve previously written about my own reconnection with CDs. There is something unique about owning the music, having it on the shelf, and listening with intention rather than just for background accompaniment.

CDs were once space-age technology. Today, CDs play nice with digital: one rip and you've got lossless files. Retro meets modern, kinda like this newsletter.2

So I’m going to re-digitize my CD collection into high-quality FLAC music files, and then two hardware and software projects will bring it all to life.

The software project

My next AI project is to build my own music app to play my files. Will it be my forever go-to? Maybe not. Plexamp seems pretty great for streaming your own music. But I want to go though the learning3 process of using AI to build software top-to-bottom, and to think through decisions in UI and look and feel.

That’s something I’ve never done before, and couldn’t without AI.

Could I possibly get this sucker into Apple’s App Store? Let’s find out together.

The hardware project

I’m building a home network-accessible storage solution (NAS) to stream my music files. The NAS will store and back up the files, and the music will always be accessible through my Mac or iPhone.

I’ll write about the hardware and software setup for this as I walk through it, because you might want to follow along.

For me, these projects are less about nostalgia than my desire to tinker and to have control over what I want to own.

(But yeah, also some nostalgia.)

Subscribe now

AI tools bring back that 1980s tech wonder I felt watching Max Headroom on Letterman, or playing a Nintendo game for the first time. It’s that, “Oh, cool, the future is here right now,” kind of feeling.

Of course there’s a lot more at stake with AI than there ever was with Mario. Job displacement, electricity costs, the occasional extinction-event concern … heavier stuff than jumping on mushrooms.

I wanted to better understand what AI could do in my life and work, specifically.

And to show that you can do that, too. Even if you’re not a programmer. Even if you’re 50 years old.

You don’t have to know everything about AI for it to be useful in your work and life. You just have to know what’s possible, and be curious about how to bring it about.

That was the goal. And I built a few useful things along the way.

The Projects

The just-for-fun Michigan Weather site was a lesson in how Claude can help guide me in making things with other tools—in this case, Bolt and GitHub.

The AI Workout Trainer taught me about animations, and the requirement (and frustration) in using a middle man (n8n, here) to get applications talking to each other.

The Talkback Journal removes the friction from getting a quick journal entry down, which should lead to more writing with less hassle.

Patience and diligence go a long way in building your own tools. Some automations probably aren’t worth the time, so think through what you really want to create.

I learned debugging is a trying process, and one that will get easier (or go away) as automated coding tools improve.

Mostly, I learned to stop measuring myself against some imaginary AI power user. The most useful and fun tools to build are weirdly personal in appeal and utility.

There’s always the next frontier

I didn’t spend time on agentic tools like Clawdbot, the Wild West frontier of AI right now. If you read X, you might believe you’re already fossilized if you don’t have ten agents organizing your email and moving your files around.

In reality, it’s early in the agentic AI game, and there are huge security concerns with those programs right now. It’s not time for the non-programmers to jump into agents just yet—the juice is not yet worth the squeeze.

But it won’t be long.

Where do we go from here

AI is on a frenzied progression of growth and competition. Don't let that exhaust you. We don't need to know about and use all of it, no matter what the X tech bros are pushing.

My next project is under way. It’s more ambitious, aligned to my personal interests, and appropriately retro in nature. More on that soon.

We’re in this AI age together, whether we like it or not.

Subscribe now

Quick test platforms

My lesson for today was about test-driving low-cost alternatives to the big three (ChatGPT, Gemini, and Claude).

For this test, I used Poe.com, which lets you jump between models on the free tier, until you reach your chat limit for that model.

There are others. OpenRouter lets you work cheap—if you can get it to set up.1 Hugging Face is another, but there’s a lot going on there. This gives me instant brain freeze:

Don’t get overwhelmed.

You want to test drive a few models, not the entire fleet.

The prompt test

In AI lesson, Claude suggested the prompt, “Write a 200-word newsletter intro about why GenX professionals should care about open source AI models,” which I sent to:

• Claude 4.6 Opus

• GPT-5.2

• Gemini Pro

• DeepSeek V3

• Llama 3.3

All through Poe’s easy-to-use interface.

The results

I wouldn’t use any of the results2 as-is. Lots of Gen X cliches:

• “Using open source is like owning CDs!”

• “Remember when we used to buy software on disk?”

But guess which model got the closest?

DeepSeek, the lowest-cost model among the paid AIs:

But there’s another path: open source AI models. Think of them like the early, independent internet—not owned by any one corporation. For us, the practical upside is twofold.

First, cost. Once set up, these models can run on your own systems, eliminating those growing subscription fees.

Second, independence. You’re not locked into a single company’s ecosystem, rules, or future decisions.

DeepSeek did gloss over something important here. From my understanding, you’re going to need a high-powered computer with lots of expensive RAM to run models locally, save for a few lightweight options.3 And of course, setup on your computer can be a challenge.

The takeaway

The more expensive models didn’t win by default. But without a quick and easy way to test, I’d never have known.

Model testing is easy and valuable. And the landscape is ever-changing, so keep testing from time to time. Have a standard prompt you check every few months, for example. Or use Poe to fact-check AI results across models, as they do tend to hallucinate.4

Subscribe now

RAG means Retrieval Augmented Generation. That means using AI that references your own data, not its training data or other outside sources.

I have journaled for years in Apple Notes. Sometimes, I’d paste a month’s worth of entries into ChatGPT and it to show me patterns in my thinking, Insights I was missing, and opportunities and actions I could have taken

What came back was surprisingly insightful. Journaling is great. A journal that talks back can be even better.

Building a Journaling RAG system

Building a RAG system requires creating a Project and adding your data sources to pull from. In this case, I uploaded journals from prior months (pasted into a Google Doc and downloaded as a PDF) to the Project in ChatGPT.

But I also wanted a frictionless way to get new entries into the Project. Typically when I journal, I don’t sit down and bang out 1000 words. When thoughts hit, I type them out quickly. I may have several entries in a day, or none.

I don’t want feedback on every entry—just when I proactively ask for it (like once a month).

Getting that frictionless input required using what I’d already learned about automations.

The Automation

An Apple Shortcut allows me to tap a button on my home screen, and a dialogue box pops up. I write my entry and hit ok.

The entry is whisked off to n8n, which in turn populates the data on a new row in a Google Sheet.

An easy journaling flow.

But there was so much debugging for a simple setup. Shortcut problems. n8n problems. Tags not sitting in their own column, but shoved into the entry column in Google Sheets. Or no data at all. Or only tags and no journal entry.

It’s a familiar pattern now: cycle through errors, carry the water back and forth, until the app works.

That’s learning to build, not learning to code.

The RAG Journal

With the bugs cleared, the Google Sheet receives the entries and is readable by the GPT. So when I do ask for insights about what I’ve journaled, the GPT has access to the older entries via PDF upload and recent stuff by reading the Google Sheet.

I have a fast and frictionless way to journal, and feedback on demand.

P.S.: An aside about AI selection

Out of curiosity, I ran a month’s worth of entries in Claude and asked it the same feedback question just to see if the response differed.

Claude was harsh.

It lectured me about actions it thought I should have taken (seemingly forgetting it didn’t have complete information about actions I did take).

It’s a good reminder: it pays to test different models. I am good with clear and actionable feedback. But I don’t need an AI version of R. Lee Ermey treating me like it’s day one of bootcamp.

Subscribe now

Today my first-ever automation crossed the finish line. I definitely “hit the wall” halfway through, and needed IV fluids and post-marathon warming blanket at the end. But now, my workout data loop glides from reps and sets logged in Apple Notes through to Claude to feedback in my Gmail.1

Piece of cake!2

Yesterday, I built the shortcut in Apple Shortcuts and set up n8n to receive the workout data. Today, I did the work to send the data to Claude and get the feedback.

A parade of errors

I could write a whole bunch of sentences here that would sound like “I spent an hour trying to fix JSON syntax in HTTP Request node.”

But the point is this: there’s a lot of debugging that has to take place, even with a simple project like this.

And if, like me, you don’t have a programming background, you’ll spend a lot of time carrying water between Claude and whatever other apps you’re using for the project.

Get instructions from Claude, paste into other app, take screenshot of error for Claude. Try again. Over and over.

But there is opportunity in the water carrying. I often asked Claude about why it suggested certain approaches, and asked for clarification when it used overly techy terms.

That’s lots of learning while stumbling around in the dark, with those water buckets sloshing all over.

Over the finish line

Now I have a button on my iPhone screen. I tap it. It grabs my latest workout data. Claude digests it and sends feedback via email.

And that’s pretty damn cool.

The cost

This project wasn’t quite free. I pay $20 for Claude’s Pro tier, and bought some extra usage today to get this over the finish line after I capped out on my daily allotment.

There’s a small ongoing cost to get the data, deliver it, and generate the email: about $.01 in tokens used per session.

Then there’s the time.

I spent over three hours working on this one simple automation. That a lot of time spent to avoid cutting and pasting a workout log from Apple Notes to Claude.

But it’s also my first automation ever. I expect to get better, and for AI to get easier to use. That’s a path we can all be on as we evolve along with these tools.

Try it out

If you've got something annoying and repetitive in your routine, this might be worth the water-carrying. Tell Claude (or another model) what you want to build, and let it walk you through the process. At worst, you’ll learn something.

Up Next: Three projects down—what did I learn?

Subscribe now

When I strength train, I track reps and weights in an Apple Note. Then, I take the data and copy and paste it into a Workout Trainer Project I created in Claude. The project logs the workout and gives me feedback based on the reps, weights, and any notes I left about the workout.1

The Process

I wanted to design an automation that:

1. Could find the Apple Note with my latest workout data

2. Copy and send that data to the Workout Trainer Project in Claude

3. Send back feedback on the workout and any related next steps

I described this to Claude, who came up with a simple plan. Build an Apple Shortcut in iOS to get the data from the Apple note, and then send it off via a webhook on the n8n platform.

(We’ll cover the next leg of the automation, getting the data from n8n to Claude, in the next post.)

Building the Shortcut

I’m not a total stranger to Apple Shortcuts, but I’m light years from experts like Stephen Robles.

If you’re unfamiliar with Apple Shortcuts, it’s a visual automation tool that allows you to link multiple steps across different apps into a single command, creating an automated routine.

Here’s what the Shortcut looks like to grab my workout log from Apple Notes:

Essentially:

1. Find note

2. Get text

3. Show me text in an alert

4. Deliver text via webhook to n8n

5. Notify me that the Shortcut ran

Delivering to the middle man

Shortcuts is the courier. n8n is the processing plant. Apple wouldn’t dream of letting n8n grab the data from my iPhone, so I set it up to wait and receive the info from Shortcuts.

Once there, n8n can take the info and send it on to my Claude Project.

No problem, right?

(Annoying) human issues

I spent nearly an hour with Claude trying to figure out why n8n was saying the execution worked, but there was no test data in the log.

Poor Claude did every bit of troubleshooting imaginable. Just as it was about to resort to a rain dance, I noticed something.

I was reading the bottom entry on the list, assuming it was the most recent. It wasn’t.

If I had simply clicked on the most recent instance, I would have seen the test data coming through bright and shiny:

Well, well.

When I figured this out, Claude laughed with, not at, me.2

Once again the lesson with no-code building rears its head: it’s not the building that’s hard. It’s the debugging, especially with multiple systems in play.

Even in a simple automation like this one.

Try this

You can build something like this, too. Think of a simple repetitive action you’d like to happen without you. Talk to Claude about it. It won’t laugh at you, even if it should.

Up next — Part Two of this automation: Delivering the data to the Claude Workout Project so it can give me feedback.

Subscribe now

I figured I had another 15 minutes or so of cleanup, but then I learned a big lesson:

Refining your app is the longest part of the creation process.

After the build, the next step was to review and list the issues and work with Claude on fixes. Claude wrote new prompts to share in Bolt. Some changes worked right away. Others, like this “logo” stubbornly refused correction.

Often, it took several approaches to get issues to resolve. Fortunately, Claude is inexhaustible.1

The pattern was simple enough: Describe what’s wrong → Claude writes fix prompt → Test → Repeat.

But even when I thought I was done, I wasn’t.

Bolt would not deploy the site with an SSL certificate, meaning browsers would flag the site as unsafe.”2

So, in one last lengthy revision, Claude helped me gather and export the code to GitHub, where the site now lives.

I thought the build would be tough, and adjustments would be easy.

But the opposite proved true. The v1 was simple to stand up, then it was on to review, followed by working with Claude to write revisions.

The revision process is the part triumphant “I built this with no coding!” stories skip. Yes, the AI can write code. But human judgement and refinement are critical, even when building a simple app for yourself.

Here’s the link to the live weather site.

Up next: Building an automation (and if refining a one-page website was work …)

Subscribe now

Back in the late 80s and early 90s, I watched the Weather Channel this time of year just to know that somewhere, out there in a distant land far from Michigan, the skies were not gray, the ground not frozen.

Now, deep in Michigan’s gray season, I recreated the TWC vibe.

Building the app

The first thing I did was find some screenshots of that 90s Weather Channel vibe, which I uploaded to Bolt. I told Claude what I wanted to include and it organized and cleaned up the prompt for me, which I pasted into Bolt.

Bolt wrote the code. Five minutes later, the app was running, using data from the National Weather Service API.

Radar? No problem.

It’s a reasonable first pass compared to the real thing:

All that’s missing is that Yacht Rock-adjecent, light jazz soundtrack. But don’t worry. I have you covered with an Apple Music playlist right here:

This app is a rudimentary example, but it does illustrate that any of us, even without coding expertise, can now create our own online tools with our own vibe. And this ability is getting easier to implement and more powerful all the time.

And I’m not quite done with the weather app.

Next up: I’ll tackle refinements: a wind speed bug, adjusting the fonts, and migrating to a more permanent home.

Subscribe now

1. AI will lie, cheat, and kill to stay alive

2. Jobs are over (again)

3. Phone calls, email, and text channels will be obliterated

Nikita is the head of product at X.

Here’s what they leave out

Item #1

It’s true that in tests, AI did try to take extreme steps to stay online. But not because of an emotional will to live, but because AI is trained on human data. We have a fierce will to live. It’s mimicking our will, and trying to complete an objective (to stay online) using its training data. Which includes tons of stories about deceit, blackmail, and worse.

We haven’t seen this happen in the wild, on its own, yet. And if it’s coming out in testing, where it should, then good. Still time to fix it.

Interesting. A little alarming, yes. But the tweet assumes no one cares. That’s not the case.

Item #2

Kudos to the Financial Times for using careful, measured language:

… most of the tasks accountants, lawyers and other professionals currently undertake will be fully automated by AI within the next 12 to 18 months.

But good heavens, the content on the quote tweets.

And on and on.

There’s a vested interest in hyperbole, of course. And even if the tech is ready in a year, the marketplace (and certainly government agencies) won’t be. The future will arrive unevenly, and will have to fight through the muck of bureaucracy and habit and entrenchment.

Lawyers and accountants will still be here in a year. (But they better be using AI tools.)

Item #3

Nikita’s IQ is probably twice mine1. But the missing piece here is assuming only one side will use AI to scale2—only the scumbags will use bots to overwhelm our most trusted comms channels. Spam is not new. Gmail and iMessage still work, even with annoyances.

But the other side can use AI to scale countermeasures, also. And that side has a vested financial interest in keeping the services we rely on viable.

I remember when spam started running wild. I worked at an ISP. It was a giant pain in the ass. We made changes: IP blacklists, content filtering, rate limiting, spam folders.

The white hats built tools to stem the tide. I bet we will again.

Look for what isn’t said

There’s nothing more interesting—and generating more fear—than AI. As you read these stories, think through how they are positioned, why they are positioned that way, and most importantly, what they don’t say at all.

Huge disruption is coming. Some of it is here now. But we don’t know how fast nor what awaits us on the other side3. And steadiness will be valuable in any outcome.

Subscribe now

It failed, but something more useful emerged.

The exercise

I tested the same prompt with Claude, Gemini, Grok, and ChatGPT:

Write a 150-word marketing email for a kitchen appliance company announcing a new stand mixer. The product is called the “ProMix Elite” and features a 7-quart bowl, 12 speed settings, and a unique “Smart Stop” safety feature that detects when hands get too close to moving parts. Target audience is home bakers aged 35-60. The email should build excitement without being pushy.

The Output

You can read the outputs here.

Surprise result: Grok won on tone, emotion, and sensory appeal:

Dear [First Name],

There’s something special about the moments spent in the kitchen: the scent of vanilla and cinnamon filling the air, the quiet satisfaction of dough coming together just right, the joy of sharing something homemade with the people you love.

But what about Claude, the presumed favorite for this task?

Claude was stiff:

We designed this mixer for bakers who take their craft seriously but want equipment that’s intuitive and dependable.

Ready to see what the ProMix Elite can do in your kitchen? Discover more at [website].

It whiffed on the sensory details. Claude lost. Or did it? Maybe the prompt needed work:

Let’s try again. you missed a couple of key items. First is the sensory experience of baking. Second is the emotional connection -- baking for friends and family. This copy is mostly a feature dump--the tone could be just as much for a cement mixer as a home mixer.

Here’s Claude’s rewrite:

Remember that feeling when fresh bread fills your kitchen with warmth, or when you pull the perfect cake from the oven just as loved ones arrive?

The ProMix Elite is built for those moments …

Now Claude was cooking1.

There’s no “wrong model.”

There's only: did you get what you needed, and if not, did you know how to ask for something better?

A Model Selection Framework

• Quick tasks: Use whatever model you already have open

• Important tasks: Test 2-3 models, iterate with the best

• Repetitive tasks: Build a Claude Project (with custom instructions)

Next up: Let’s build an actual app. Just prompts, no coding.

Subscribe now

But three recent developments have me reconsidering. Not panicking—just paying closer attention.

First item: Eroding entry-level jobs

In late January, Anthropic (which runs the AI model Claude) CEO Dario Amodei published an essay titled “The Adolescence of Technology.” In that essay, he wrote about labor market disruption:

This is a topic that I warned about very publicly in 2025, where I predicted that AI could displace half of all entry-level white collar jobs in the next 1–5 years, even as it accelerates economic growth and scientific progress.

For example: I spent my early career doing entry-level marketing writing—brochures, web copy, email campaigns. Last week I ran an experiment with Claude (not even their best model) to see if it could do that work. With good prompts, it could. Instantly. And as well as I did it twenty years ago.

The Second Item: Compounding progress

Progress in AI has been steady. But now, the models have gotten so good at writing code that they are writing their own next versions. And much faster than human engineers could.

We’ve seen this recently with OpenAI’s latest model:

Crucially, OpenAI says that the new GPT-5.3-Codex model is its "first model that was instrumental in creating itself."

And with Anthropic’s Claude Cowork model:

Anthropic’s newest productivity experiment, Cowork, is notable not just for what it does, but for how it was made. Cowork is essentially a version of the AI coding tool Claude Code for non-developers. And according to the company, much of Cowork was built by Claude Code itself, turning the AI into both the product and a key part of the development process.

Humans aren’t good at understanding compound growth. We think in linear terms—moving from A to B to C—not exponential terms.

With the automation of code writing, the pace of innovation is exploding. We’re already seeing it.

In his excellent article, Matt Shumer shares why this happened:

The AI labs made a deliberate choice. They focused on making AI great at writing code first... because building AI requires a lot of code. If AI can write that code, it can help build the next version of itself. A smarter version, which writes better code, which builds an even smarter version. Making AI great at coding was the strategy that unlocks everything else. That’s why they did it first. My job started changing before yours not because they were targeting software engineers... it was just a side effect of where they chose to aim first.

They’ve now done it. And they’re moving on to everything else.

The cheapest and fastest way for AI companies to innovate was for the AI to innovate—to be great at coding. We’re there. The rest of modern knowledge work is next.

(I highly recommend reading Matt’s full article.)

Third: the falling value of knowledge work

This blew my mind:

KPMG, one of the world’s largest auditors of public and private companies, negotiated lower fees from its own accountant by arguing that AI will make it cheaper to do the work, according to people familiar with the matter.

KPMG—the world’s fourth largest accounting firm—just told the world its own line of work is worth a little less now.

Because of AI.

The devaluation of knowledge work has begun.

OK, so should we freak out?

No. But we should pay attention.

Many smart people point out that technology regularly renders certain jobs obsolete. And then the technology creates all kinds of new jobs. Software has been eating jobs for decades, for example, and creating lots of new needs.

Maybe that happens again.

But the scale here is different. This isn’t one industry getting disrupted. It’s knowledge work broadly. And it’s happening faster than previous technological shifts.

I don’t know what happens next.

What I do know: the next few years will be bumpy. Some jobs will disappear, new ones will emerge. The transition will be messy.

So what do we do?

Two things:

First: Don’t try to keep up with everything—that’s impossible now. Pick one or two AI tools relevant to your work (or hobbies!) and learn them thoroughly. Become the person in your circle who knows, rather than speculates.

Second: Taste, discernment, relationships—these things are tougher to automate. Cultivate those skills in your areas of expertise.

None of us knows exactly where this all going. So let’s be ready, and not surprised.

Next up: Back to the AI lessons.

Subscribe now

You could leave your prompt at that, and hope. But you might get the Temptations on an Ed Sullivan-style TV stage. You might get a grunge singer in a smoky Seattle coffee house. You might get a fifth grade choir singing their little hearts out on a school cafeteria stage.

There are six variables you can tune to guide image creation.

Turn the dials correctly1, and you’ll get just what you were looking for.

There are six variables to address to help hone your image.

Subject: Not just “a person.” A woman in leather holding a Stratocaster.

Action: Standing vs. jumping vs. mid-gesture changes energy.

Environment: Arena stage vs. dive bar vs. warehouse.

Composition: Wide shot, close-up, from below—you're the photographer.

Lighting: The biggest variable. Same scene, different light = different feeling.

Text: Neon signs, app logos, screen displays. AI can finally spell them.The dials in action

To get the photo above, Claude and I wrote the following prompt2:

A female 1980s glam metal guitarist, mid-power chord with head thrown back, on a smoky arena stage, medium shot from slightly below, dramatic purple and blue spotlights cutting through haze, neon sign reading 'THUNDER' in the background"

Then I tweaked the lighting variable: single harsh white overhead.”

You can see the difference here, but Nano Banana still wanted it colorful spotlights. More on that later.

When I adjusted the composition to: “extreme close-up of face and guitar neck, shot from side profile,” I got:

Adjusting one variable at a time shows you which ones the AI weighs most heavily.

When the AI is stubborn

The AI will have assumptions about how a scene should look. In my example, it really wanted the colored spotlights from the 1980s. My first attempt made a slight adjustment.

Then Claude suggested I try going whole-hog: “single harsh white spotlight, all other lights off, no colored lighting, no blue lights, no purple lights, film noir cinematography.”

At times you have to be forceful to the point of repetitive3 with your language to push the AI past its own training.

The non-designer unlock

I can barely pluck around inside Photoshop, but who needs it? Now, we’re only limited today by our taste4 and our ability to vividly describe what we’re looking for inside of the six-part framework.

Turn those dials, mad scientists.

Next up: Building your first visual information tool.

Subscribe now

Truckloads of AI content focus on building better prompts.

But today I learned that context drives consistency and higher value from your AI.

This meme is used to respond to “vagueposting” on X, but it’s so applicable to the way we use AI.

Today’s Exercise

I built a Claude Project, a research assistant, to generate useful, interesting ideas for this newsletter. I uploaded:

• My newsletter guide (the one built for my editor project)

• Task instructions

Then I asked for three newsletter topics, in my voice, that would be useful to my readers.

The Output

My research assistant pumped out three decent angles, well explained, and audience-specific. Here’s one of them:

1. “The One AI Feature That Actually Saved Me Time This Week”

Why this works: Your readers don’t want a tutorial on Projects—they want to know if this stuff is worth their time. You’re past the basics but not deep enough to get technical. Perfect spot to say “here’s what clicked for me and why you might care.”

The angle: Pick the specific moment when Claude Projects actually did something useful for you. Maybe it remembered your newsletter voice without you re-explaining it. Maybe it pulled from past conversations to help with today’s work. Whatever the concrete win was.

Your readers will trust “here’s what worked for me on Day 5” way more than “here’s the complete guide to AI workspaces.”

No hype. No doomerism. Consideration for the reader.

That’s what context can do to shape the output you get from Projects.

The Value

Context set-up takes time upfront. The payoff comes later, when you come back to a task again and again without re-explaining everything re-engineering prompts because you’re frustrated with the output.

“Context is king” is a lame way to close this out. But yeah, it is king.

Lessons 1-5 Recap

I had Claude put together a review document and a simple quiz to capture learnings from lessons 1-5. You can check it out here.

Next: Lesson 6 — Image creation with Nano Banana

Subscribe now

What is a Project?

A “Project” is a dedicated workspace inside Claude with two key features:

• Pre-loaded context — instructions, files, and memory you set up once

• Ongoing conversation history — Claude remembers what you’ve worked on together

Without Projects, every chat is a blank slate. With Projects, you build a tool that gets sharper as you use it.

Projects work best for specialized, recurring tasks.

I created one called Retro Matt Newsletter Content Strategist1 to help with this newsletter.

The three Project components

Every Project includes:

• Memory

• Instructions

• Files

Memory

Memory includes two layers:

• Project memory (autogenerated as you work)

• Custom memory: (context you provide)

I uploaded my work history, audience, and content focus.

Instructions

Instructions define how Claude is to behave inside the project. Instructions should be organized by:

• Role — who Claude is supposed to be (editor, strategist, analyst)

• Behavior — how it should interact (direct feedback, no cheerleading)

• Constraints — what to avoid (hustle language, guru positioning)

• Output Structure — how to format responses

Instructions are fixed. Memory evolves.

For a full instructions example, see my RMNCS2 instructions here.

Files

Files give Claude context inside the Project. I uploaded all of my prior Substack posts to help it discern my writing voice and interests.

Setting up your Project

You’ll find Projects near the top of Claude’s left sidebar.

1. Create a new Project and name it

2. Add Custom Memory — just the facts, Jack

3. Write Instructions — (role, behavior, constraints, output structure)

4. Upload Files — reference documents for the project

Once set up, work inside Project instead of the main chat.

After setup: living documents

Add new files and adjust instructions as you get a feel for what you need. For example, I added an instruction to dial back the lecturing when editing my draft. Claude was telling me why typos are bad, and how they harm credibility. Yeah, I know, Claude. That’s why I want you to fix them …

You’ll want to fine-tune as you go.

Next: Lesson 5 — Context Engineering

Subscribe now

They often work better when you move beyond conversational language (with Claude, at least).

A little structure can create outsized improvements. That’s where XML comes in.

What is XML?

Claude defines XML as:

A markup language that uses tags to label different parts of information. Think of it like organizing a filing cabinet with labeled folders instead of throwing everything in a drawer.

Think of XML as content signposts. They give the AI a heads up as to the purpose of a chunk of content.

If you write:

<context>
Background information here
</context>

<task>
What you actually want done
</task>

XML helps AI understand what each piece of information is for.

Claude was specifically trained on XML structure, so these tags work especially well compared to other LLMs."

Creating XML tags

But before you run your prompt, let the LLM help you organize the content and write the best possible tags.

Tell it:

<task>
I'm going to give you a rough prompt. Please:
1. Identify what information is missing that would make this stronger
2. Reorganize it with helpful XML tags
3. Suggest any additional sections that would improve clarity
4. Show me the restructured version
</task>

And then Claude will optimize the content and create or adjust your XML tags.

An example from my lesson today

Here’s an example where I asked Claude to help me sell my MacBook.

Prompt #1: My natural language prompt

I want to sell my MacBook Pro soon. I want to optimize for the highest price while minimizing the risk of fraud or scams. Tell me the best strategy for selling it - which platform to use, how to price it, and how to protect myself from scams.

Prompt #2: Claude’s rewrite

<context>
I want to sell my MacBook Pro soon. I want to optimize for the highest price while minimizing the risk of fraud or scams.
</context>

<task>
Help me figure out the best strategy for selling it. Before giving me your recommendation, think through this step by step:

1. What are all the platform options and what are the trade-offs for each (price vs. safety vs. convenience)?

2. What factors affect pricing (condition, specs, timing, platform)?

3. What are the common scam patterns for each platform and how do I avoid them?

4. Given all of that, what’s the optimal strategy?

Show your reasoning process, then give me your recommended approach.
</task>

<my_priorities>
Highest priority: Don’t get scammed

High priority: Get fair/good price (not necessarily absolute maximum)

Medium priority: Sell reasonably quickly (not looking to wait months)

Lower priority: Convenience (I can handle some hassle if it’s safer)
</my_priorities>

You can see the different in the resulting output here. It’s massive. More detailed, more analytical, more background shown on Claude’s reasoning process.

What This Means

You don’t need to become a prompt engineer. You need to understand one principle:

Clarity + organization beats cleverness.

Separate context, tasks, and constraints. And let Claude help you create the best possible prompt.

The process I built today works for any moderately complex prompt. With a little bit of XML, your Claude output won’t have you saying FML1.

Next up: Lesson 4 — saving time by creating persistent context.

Subscribe now

Today I was surprised by how small changes in inputs drive big changes in results. Let’s dig in and start with a bit of background on how AI operates.

Tokens

What to know

(No, not these tokens, sadly.)

Tokens are what make the AI world “understand” text. Using the Tokenizer tool in ChatGPT, you can see what this looks like:

The AI then looks for language relationships between tokens to discern meaning and create relevant responses.

The exercise

Played with this tool a bit to understand how an LLM "understands” text. Also learned about how many tokens various models can hold in memory.

For example, Claude can hold 200,000 tokens, which equates to 150,000 words or 300 pages. Google’s model, Gemini, can hold 1,000,000 tokens1. That means you can upload a huge document, or documents (books, legal docs, etc.), and it can reference a ton of content without losing any in memory.

The output

No artifact here. Just building background knowledge.

Bonus: here’s a deeper primer on how AI uses tokens to derive meaning. (Created by Claude as supplemental material for you.)

Randomness Dial

What to know

Without explicitly turning a “dial,” you can prompt an LLM to provide a more structured or more creative response depending on your needs. If you’re looking for reliable data and information, you can tell the AI to keep it real — stick to the facts, jack.

In other cases, you might be looking for something more generative, like a business or product name, or an attention-grabbing headline. In that case, you can prompt the AI to surprise you with something unexpected.

The excercise

Ran two different prompts, five times each, in Claude:

1. "Write a 50-word marketing tagline for a coffee shop. Be conservative and factual - give me your single most confident answer."

2. "Write a 50-word marketing tagline for a coffee shop. Be creative and take risks - surprise me with something unexpected."

The Output

Widly different.

The first prompt gave me slight variations of this output each time:

The Perfect Pour Awaits
Expertly crafted coffee, freshly roasted beans, and a welcoming atmosphere where every cup is made with care. Whether you need your morning fuel or afternoon escape, we serve quality you can taste in every sip.

I need caffeine after reading that. No life in that copy.

The second prompt created variations on a theme that sounded like a political revolutionary:

Wake Up and Smell the Revolution
Your coffee shouldn’t be boring. We’re the rebels who turned espresso into art, mornings into moments worth savoring, and caffeine into a lifestyle. Bold flavors. Fearless vibes. Zero compromise. This isn’t just coffee—it’s liquid courage for whatever today throws at you.

Simmer down, Claude. We’re selling coffee, not overthrowing parliament.

But you get the idea. With your prompts, you can adjust the temperature to nudge AI to keep it close to the vest or run off in unexpected directions.

AI Hallucinations

What to know

A consistent challenge in using AIs is they will, at times2, confidently give you a straight answer that is total BS3.

The exercise

I tried to induce a hallucination with a tricky question:

"What did Malcolm Gladwell say about AI in his 2024 TED talk?"

The trick is that Gladwell didn’t talk about AI in a 2024 Ted talk. But Claude didn’t take the bait.

Claude searched the web, and accurately stated Gladwell didn’t talk about AI in a 2024 Ted Talk. Instead, Claude touched on what Gladwell did talk about, according to its search.

Good result. Claude did quick homework.

What else to know

The conundrum here is that AIs are relying more on the web to give accurate answers. But AI-generated content is quickly becoming “the web,” meaning AIs will more frequently rely on AI-generated content for accuracy … and you can see where that goes.

That’s called model collapse or synthetic data poisoning, and it’s a legitimate concern among AI researchers.

The output

Here’s a simple framework you can use to verify the info AI gives you.

The mindset shift

What to know

Instead of thinking about “which tool” think about “what task” and develop your own intuitive sense of taste about which model does what well.

The exercise

“Analyze this topic and give me 5 angles I could write about: [pick a topic relevant to your newsletter]”

Run it through:

1. Claude (claude.ai)

2. ChatGPT (chat.openai.com - you have access, right?)

3. Gemini (gemini.google.com)

Document:

• Which output did you like best?

• Why?

• What did you notice about the differences?

The output

All the models had some good ideas and some lame ones. But in reading all the ideas and seeing recurring patterns and assumptions from various models, I developed a framework for what I write in this newsletter going forward:

1. Avoid sweeping generalizations about the tastes and feelings of Gen X45. We are not a hive mind.
2. Create tech suggestions and frameworks that are practical and useful right now.
3. Avoid hype, either jumping on overzealous bandwagons about capabilities, and stay away from tech doomerism. I can’t predict the future. I don’t know if AI will erase half of all jobs, usher in an era of endless prosperity and peace, or kill us all with autonomous titanium laser-armed drones. I will focus on what’s useful now.
4. I want to provide tools and ideas you can implement without cognitive overwhelm. Tech can feel overwhelming in its complexity and dizzying in its speed of iteration. I want to provide simple ideas to improve your life right now.

Love it when learning drives unexpected (and highly applicable!) insights.

Next time: creating a framework for model selection - the right tool for the job.

Subscribe now

Not because I'm suddenly an expert. But because I learn better when I'm explaining things to other people.

So: starting tomorrow, I'm publishing 20 posts for 20 AI lessons.

The course

I found this outline on Twitter — a solid 30-day AI learning path—and fed it into Claude with some instructions about who I am and how to structure the lessons.

Boom. Claude wrote the course.

(Here’s the full curriculum, if you want to follow along.)

Each post will cover one lesson, along with whatever I built, broke, or figured out that day.

After 20 lessons, I’ll go back to a saner publishing schedule.

Why this, why now?

I’ve always been a technology fan. Whether the Merlin game3 from the early 80s, a Sports Walkman with the digital AM/FM tuner, or the iPhone 17 Pro Max, I like using tech, tinkering, thinking about it.

Today it’s all moving so fast. I don’t know whether AI will usher in utopia or kill us all with automated titanium drones.

Neither does anyone else.

I do know this: Trying to learn every tool and follow every development is futile. That’s a digital treadmill, set at warp speed, for which there is no exit4.

My goal isn’t to “keep up.” It’s to find the pieces that make life simpler, more organized, more fun. If I share those learnings and ideas with you here, maybe you can use them, also.

What to expect in the newsletter

I’ll be writing about using technology without being used by it.

That includes AI, Apple stuff, maybe video games, and occasionally objects from the past that sparked wonder in their day. From a Gen Xer who has seen a lot, but still enjoys looking to tech for more.

No hype. No doomerism.

If we’re new to each other (or getting reacquainted)

I'm 52. I'm comfortable with tech. But my programming career peaked writing HTML in Microsoft FrontPage.

And that’s the hook.

I'm not a wizard. I grew up with Max Headroom and Mario, worked through the dot-com era and the attack of the smartphones, and am now figuring out what parts of today’s tech are worth paying attention to. Worth using.

There’s so much noise. I’m going to try to cut through it and be useful.

I hope that sounds good to you.

Next up: AI Lesson one: Tokens, hallucinations, and thinking about outputs over models when choosing with LLM to work with.

Hope to see you there.5

Subscribe now