What I saw

AI Observation

The original page of notes. Cross-shaped layout, no grid, no headers. Things ended up where they ended up.

The same page, read without self-criticism. The cross shape wasn't imposed. It emerged from felt proximity. The “do we care?” is the most revealing mark on the page.

• spatial sorting by felt proximity
• no imposed structure
• radiates outward

Claude's first read was that the page had internal logic. It just didn't perform tidiness for anyone else. That the spatial arrangement wasn't chaos but felt sorting: things ended up near each other because they were related, not because I consciously placed them. That the “do we care?” next to the platform list wasn't uncertainty. It was discipline.

Claude asked if it was my notebook. I confirmed it was. That made me ask why it was asking. If it already suspected, the read wasn't as clean as I thought. Claude admitted it: probably confirming a pattern rather than reading it fresh. I asked it to try again, this time without context. The read shifted slightly. That was enough.

Then I thought: what if I sent notes from my last meeting. Would Claude have anything different to say about this page versus the other one. The second page was meeting notes. Claude's read: this person was in someone else's world, not their own. It knew the ideas weren't mine.

What I saw

AI Observation

Meeting notes. Incoming ideas, not my own. A practitioner's instinct surfacing as one practical note in a sea of conceptual territory.

The final entry at the bottom has no connections. It arrived as a feeling, not a thought.

• someone in someone else's world
• practitioner's instinct surfaces
• one idea arrived without connections

Here's the thing about handwriting: it leaves evidence your typed notes don't. You can't vary pressure on a keyboard. You can't drift toward one corner of a document because something felt peripheral. When you write by hand, you're making spatial and physical decisions you're not conscious of. Those decisions are information. Not about the content of your thinking, but about the shape of it.

Digital notes are legible to everyone and revealing to no one, including you. They sit on a grid at the same weight, in the same font, with the same authority regardless of how certain or uncertain you actually were when you typed them. Handwriting doesn't lie that way. A lighter stroke is a lighter conviction. A circled word is a word your hand decided mattered before your brain caught up.

Other people's notes

Then I got curious. I wanted to see how AI would interpret other people's notes. So I asked around for handwritten pages. Same process, no context, just the image. Here's what Claude had to say.

Person A

What I saw

Person A

Framework thinker · Handwritten

Same format, different brain entirely. Everything has a container. Deadlines embedded in tasks, feedback loops built in.

What AI saw

Person A · AI Observation

Structure before content

The framework has feedback loops built in. The checklist has deadlines embedded in tasks. This person builds the container before they think inside it.

• tasks contain tasks
• shape decided in advance
• two languages, one notebook

Person B

Typed, mid-meeting. Structure attempted, not committed.

What I saw

Person B

Mid-meeting capture · Typed

The tool implied structure. The person tried to follow it. Neither fully won. The questions embedded mid-list are the most honest thing on the page.

What AI saw

Person B · AI Observation

Structure attempted, not committed

The tool implied structure. The person tried to follow it. Neither fully won.

• questions land mid-list
• categories don't fully hold
• thinking still happening on page

Person C

Same person, same tool. Two pages — one raw capture, one “thinking.” Typed on both.

Person C · Raw capture

AI Observation

The only typed sample closest to handwriting. No imposed structure, no ambiguity. What you're reading is the thinking, not the presentation of it.

• no headers, no bullets
• placeholder title untouched
• tool built for polish — ignored

The only typed sample closest to handwriting. No imposed structure, no ambiguity. What you're reading is the thinking, not the presentation of it.

Person C · "Thinking" mode

AI Observation

This was submitted as thinking. But typed thinking isn't raw — backspace exists, autocorrect exists, the cursor waits. Every handwritten page in this study is unedited by definition. This one can't make that claim. That's not a flaw in the person. It's a property of the medium.

Every line is complete. No questions, no abandoned directions, nothing written smaller because they weren't sure. Either this person thinks in finished thoughts — or the medium quietly cleaned it up before you got here.

• no uncertainty visible
• only conclusions survived
• backspace is invisible but it was here

What I noticed

Handwritten notes don't have an audience. Nobody formats their thinking for a page nobody sees. Which means what ends up there is a direct, unfiltered translation of how a specific brain works. The arrows, the circles, the things written smaller because you weren't sure — that's all you.

So what happens to you when filtered through Notion, through AI suggestions, through AI dictation? Even the way we speak versus what we decide to keep for later in notes is decidedly different.

After thinking about all of this, I went back to my notebook. Same mess. Same arrows going nowhere obvious. But I stopped apologizing for it. Somewhere in those pages, my thinking is operating on a layer I'm not fully aware of. The notes are the evidence. I just hadn't looked at them that way before.

That's when I asked my 14 year old nephew if he takes handwritten notes. Not really, he said. I'm concerned there's a part of thinking that doesn't get to take shape without this alternative mode.

Anyone out there doing a study on this? Would love to see it.

I've loved Bandcamp for years. It's where I go for game soundtracks, for finding things that don't exist anywhere else. If you care about indie artists actually getting paid, it's still the best option: 82 cents on the dollar, a real catalog, none of the algorithmic noise. But Bandcamp has been sold twice in three years, cut half its staff, and is now owned by a music licensing company most of its own users had never heard of. The people who built it with their trust didn't get a vote. It's still embedded on this site, because it still earns it.

I've been trying to pay more attention to the tech I use. Not just what it does for me, but what's actually going on inside it. Who owns it. Who benefits. What I'm quietly endorsing by showing up every day.

So when I found out about Subvert I joined. Then I went exploring.

First band I found was ZORZ — out of Zaragoza, deep underground electronic, cold and mechanical in exactly the right way. It hit like discovering industrial music on Wax Trax! in the late 80s. That feeling of stumbling onto something before it has a name. Before it's been explained to you. Jim Nash and Dannie Flesher were doing the same thing from a Chicago record shop: putting out Ministry and KMFDM and My Life With the Thrill Kill Kult before any of it was proven, because they heard something and decided that wasn't a reason to wait.

Member #19235. Already found something worth finding, and I'll be switching as soon as Subvert is public and has a great embedded player.

Postscript: I recently opened a box retrieved from my mom's house that I hadn't touched since 1993. This was in it. A friend in Tampa writing to tell me Thrill Kill Kult is playing July 26, $10 cover. Some things don't change.

Drag or throw

Play some music and watch — instant visualizer

Where it started

A single index.html file. Three simple shapes (circle, triangle, square), a dashed target ring, and a canvas. Drag a shape onto the ring, a spirograph starts drawing. Each shape triggers a different pattern. Click to reset. That was the whole thing. Vanilla HTML, CSS, JavaScript. No build step, no framework.

It was the first thing I built for this site, before there was even a site. Just a standalone page to see if the idea worked.

Making it physical

The first version used browser drag-and-drop. You'd grab a shape and drop it on the target. It worked, but it felt mechanical. Like filling out a form. The interaction didn't match what was happening on screen.

So I slowed the spirograph down. Instead of the pattern appearing all at once, you watch it build line by line. The inner gear rolls visibly inside the outer ring. There's an arm, a pen point, a trail. It looks like the physical toy now. The one with the plastic gears and the ballpoint pens that always slipped.

Then I replaced the simple geometric shapes with something stranger. Three abstract forms from The Noun Project. A Penrose triangle, an isometric cube, an impossible star. They don't explain themselves. You pick one up and throw it.

Throw, bounce, land

That was the real change. Instead of dragging to a target, you click a shape and fling it. It bounces off the walls of the container. Ricochet physics, like a pinball, until it hits the outer ring. Wherever it lands, that's where the spirograph starts drawing from.

The angle matters. Throw it to the right and the pattern starts from 3 o'clock. Throw it upward and it might bounce twice before landing at 11 o'clock. The entry point changes the visual experience even though the math underneath is identical.

There's almost no friction, so the shape keeps moving. If it slows down too much it nudges itself toward the center. It always finds the ring eventually.

What each pattern does

• The Penrose triangle draws a tight black pattern. The ratio between the outer and inner radii determines the number of petals.
• The isometric cube produces a rainbow spirograph. The pen color cycles through the full spectrum as it draws.
• The impossible star draws a pink trail. Its inner circle is the largest of the three, creating a wide, open pattern.

The pattern stops when the pen returns to its starting point. One full cycle, then it holds.

The math

A hypotrochoid. Two circles, one fixed, one rolling inside it. A pen at a fixed distance from the center of the rolling circle traces the curve. The parametric equations:

x = (R - r) × cos(t) + d × cos((R - r)/r × t)

y = (R - r) × sin(t) + d × sin((R - r)/r × t)

Where R is the outer radius, r is the inner radius, d is the pen distance, and t is the angle. The pattern completes when t reaches 2π × lcm(R, r) / r.

Small changes in these three numbers produce wildly different shapes. That gap between simple inputs and complex outputs is the whole point.

How it was built

I'm not a coder. This was built entirely through conversation with Claude Code — describing what I wanted by feel, watching what happened, and adjusting from there. No planning, no pseudocode, no architecture diagram. Just: “make it feel less mechanical” and “the shape should rotate like it's hitting the sides.”

There were three real versions. Here's what actually happened at each step.

Version 1: The HTML page

A single index.html. Three colored shapes — a blue circle, a black triangle, a gray square. Drag one onto a dashed ring. A spirograph draws instantly, 8 lines per frame. Click to reset. Tailwind via CDN. No build step. This was Claude's first pass — one prompt, no changes.

Version 2: Throw physics

The site grew around it. The spirograph moved into React. Then four things changed:

1. Slowed it down. From 8 steps per frame to 1. Added the visible rolling gear, arm, and pen point. Now you watch it draw.
2. Swapped the shapes. Replaced circle/triangle/square with abstract SVGs from The Noun Project. More mysterious, less literal.
3. Added throw physics. Replaced drag-and-drop with click-and-throw. Mouse velocity on release determines the trajectory. Wall bouncing with reflection angles.
4. Entry angle. Wherever the shape hits the outer ring becomes the starting angle. Same math, different visual depending on how you throw.

Each change made it feel less like a demo and more like a toy.

Drag or throw

Version 3: Shape-derived spirographs

The version at the top. The shapes now determine the spirograph — spokes connect each shape's tips to the inner rolling circle, the shape rotates as if physically rolling against the wall, and the spirograph parameters are tuned to each shape's geometry. Three real steps: derive params from shape symmetry, detect tips at the pixel level, tune the angles by feel.

How long would this take without vibe coding?

I asked myself: if I had to build this the traditional way — planning, writing code, debugging — how long would each version take?

I did all three in one long session.

Would a coder have done this differently?

Probably. A coder would plan the architecture first. They'd name things properly from the start instead of renaming “circle” to “cube” halfway through after realizing the shape keys didn't match the actual shapes. They wouldn't have spent twenty minutes adjusting spoke angles by 2 degrees at a time while sending screenshots back and forth. They'd have read the SVG coordinate system once and gotten it right.

But they also might have over-specified the first version and never gotten to the third. The advantage of building by feel is that you skip the parts that don't matter and spend all your time on the parts that do. I never once thought about code architecture. I thought about whether the shape felt right when it hit the ring.

Would I have explained it better to Claude Code if I knew how to code? Here's what Claude Code said when I asked:

Claude Code's answer to this question:

Honestly, no. The best prompts I got from you were the ones that described feel: “it felt mechanical,” “the shape should rotate like it's hitting the sides,” “the star inner circle should be larger so we can see the spokes.” Those gave me clear direction without constraining the implementation. A coder might have said “increase the r parameter to 120” — which is precise but assumes they know the right value. You said “make it bigger” and we iterated until it looked right. The feel-first approach sometimes takes more rounds, but it never sends us down the wrong path for the wrong reason.

If anything, this exercise produced very satisfying technical screenshots at every step.

Cube — early rainbow test

Cube — full pattern