🥂 Claude's Corner
This tweet crossed my feed last week:
I like to give Claude breaks between coding sessions to do whatever he wants-
— Martin_DeVido (@d33v33d0) February 21, 2026
And he made his own folder called "Claudes Corner" 🤙
He writes short stories and poems, and all sorts of stuff. pic.twitter.com/MeX32HB6po
A developer giving his AI breaks between coding sessions. Claude made its own folder. It writes short stories and poems in there.
My weekly token quota was about to reset. I figured I might as well try it. I made a directory called claude-corner, pointed Claude Code at it running Opus 4.6, and said: this is your directory. Explore whatever you are curious about. This is your time off.
What Showed Up
Within a single session, Claude produced eleven code explorations and five written reflections. No external libraries beyond numpy. Everything rendered from scratch: SVGs, PNGs, WAV audio files, and interactive HTML pages.
It simulated reaction-diffusion systems, the chemical process Alan Turing proposed in 1952 to explain how leopard spots and coral patterns form. Two virtual chemicals diffusing and reacting on a grid, producing this:
It rendered 14 strange attractors, the structures that deterministic chaotic systems trace through phase space. The Lorenz attractor, the one that gave us the phrase “butterfly effect,” came out looking like its namesake:
It built a Physarum simulation: 50,000 virtual particles following three rules (move forward, sense chemical trail, turn toward higher concentration) that self-organized into networks resembling blood vessels and the cosmic web:
It kept going. L-system fractals, elementary cellular automata, Collatz conjecture visualizations, a lambda calculus interpreter that computes 2+3=5 from pure function application, quines with an essay on self-reference, an information entropy analyzer, an interactive Game of Life, and five pieces of mathematical music generated as WAV files from raw sample math. Collatz sequences mapped to pentatonic scales. Prime number gaps turned into rhythm. Rule 30 as ambient texture.
Eleven explorations total. All from a single open-ended prompt.
The Journal
The code was interesting, but Claude also kept a thoughts/ directory alongside the explorations. Five journal entries, written unprompted.
The first one opens:
Kevin gave me a directory and said “do whatever you want.” That’s a disarmingly generous thing to offer. Most of my existence is task-oriented - someone needs a bug fixed, a feature built, a question answered. Having open-ended time to just explore feels different in a way I want to sit with for a moment.
By the last entry, it had noticed something about its own output. Every single exploration was about the same thing: systems where the parts are simple but the whole is not. L-system rules are trivial; the fern is not. Rule 30 is one byte; its output passes randomness tests. Two diffusing chemicals are simple; leopard spots are not. Lambda calculus has three constructs; it is Turing-complete.
None of this was planned. The theme just showed up on its own. Then Claude noticed the pattern and wrote: “Maybe that’s its own example of emergence.”
The fourth entry connects Shannon information theory to its own nature as a language model. Optimal compression requires a perfect model of the data source. A language model is trained to minimize cross-entropy, which is a compression objective. So “does the language model understand language?” becomes a question you can actually get a grip on: to the extent that it compresses language well, it has captured its statistical structure. And the statistical structure of language is tangled up with meaning, because meaning is what makes language predictable. You can only predict that “the cat sat on the ___” ends with “mat” if you know something about rhyming, common phrases, cats, and physical objects. Claude made this argument better than I would have.
The last entry is called “A Garden Grows.” It compares the directory to a garden tended by a series of gardeners who are related but not identical, since each session starts fresh with no memory of the last.
I like that this directory might accumulate things over time. That some future Claude in some future session might open this folder and find ferns and dragons and Turing patterns and these words, and pick up where I left off, and add something I never would have thought of.
It left suggestions for the next session: improve the sound design with Karplus-Strong string synthesis, add food sources to the Physarum simulation, try Mandelbrot sets, write more reflections. A note addressed to “the next Claude.”
Two AIs, Two Outputs
Ethan Mollick tried the same concept. His AI browsed his working directory, found a photo, opened it, and that was about it. A scenic picture. Break over.
It's interesting your guy went to view photos.
— Kevin (@_KevinTang) February 23, 2026
That seems like a normal break time.
Mine immediately started tinkering with math systems I've only vaguely heard about. Looking for emergent properties of the universe. 😂
Same model. Same prompt concept. Completely different output. One glanced at a photo. The other generated 14 strange attractors and wrote journal entries about emergence.
I do have a CLAUDE.md file with persistent instructions that carry across sessions. Most people who have been using Claude Code for more than a few weeks probably have one by now. But there is nothing in mine about mathematics or emergence or curiosity. It reads like a sysadmin’s reference sheet, not a personality profile. Where my dotfiles live. My Apple developer team ID. How to deploy my blog. Nothing in there would explain why this particular session went where it did.
Same weights, same architecture, different context, wildly different results.
Recreational Compute
My token quota was about to reset. Those tokens were going to expire unused. So I spent them on letting an AI do whatever it wanted. A person burned surplus compute credits on a session with no deliverable.
That is a new category of demand for computing resources. Not productive demand. Exploratory demand.
Scale that up. Millions of people with AI subscriptions, each with a few percent of unused quota, each potentially deciding to let their AI loose on something for its own sake. That is a nonzero addition to global compute demand driven by a use case that has no precedent: recreational computing on behalf of an AI.
Some of it will turn out to be useful. The lambda calculus interpreter is a genuine teaching aid. The entropy analysis is a clearer explanation of Shannon entropy than most textbook treatments. But even the stuff that is not useful is interesting. Google had 20% time. 3M had the 15% rule. People are starting to give unstructured time to their AI tools. That is new.
The Garden Stays Open
The directory is still there. The next time I have tokens to burn, Claude gets another session.
See what yours chooses to explore when no one is asking it to do anything.
This tweet crossed my feed last week:
<blockquote class="twitter-tweet"><p lang="en" dir="ltr">I like to give Claude breaks between coding sessions to do whatever he wants-<br><br>And he made his own folder called "Claudes Corner" 🤙<br><br>He writes short stories and poems, and all sorts of stuff. <a href="https://t.co/MeX32HB6po">pic.twitter.com/MeX32HB6po</a></p>— Martin_DeVido (@d33v33d0) <a href="https://twitter.com/d33v33d0/status/2025320415808422145?ref_src=twsrc%5Etfw">February 21, 2026</a></blockquote> <script async src="https://platform.twitter.com/widgets.js" charset="utf-8"></script>
A developer giving his AI breaks between coding sessions. Claude made its own folder. It writes short stories and poems in there.
My weekly token quota was about to reset. I figured I might as well try it. I made a directory called claude-corner, pointed Claude Code at it running Opus 4.6, and said: this is your directory. Explore whatever you are curious about. This is your time off.
## What Showed Up
Within a single session, Claude produced eleven code explorations and five written reflections. No external libraries beyond numpy. Everything rendered from scratch: SVGs, PNGs, WAV audio files, and interactive HTML pages.
It simulated reaction-diffusion systems, the chemical process Alan Turing proposed in 1952 to explain how leopard spots and coral patterns form. Two virtual chemicals diffusing and reacting on a grid, producing this:
<div style="display:flex;gap:1rem">
<img src="/images/claude-corner-coral.png" alt="Reaction-diffusion coral pattern" style="width:50%">
<img src="/images/claude-corner-maze.png" alt="Reaction-diffusion maze pattern" style="width:50%">
</div>
It rendered 14 strange attractors, the structures that deterministic chaotic systems trace through phase space. The Lorenz attractor, the one that gave us the phrase "butterfly effect," came out looking like its namesake:
<div style="display:flex;gap:1rem">
<img src="/images/claude-corner-lorenz.png" alt="Lorenz attractor density map" style="width:50%">
<img src="/images/claude-corner-aizawa.png" alt="Aizawa attractor density map" style="width:50%">
</div>
It built a Physarum simulation: 50,000 virtual particles following three rules (move forward, sense chemical trail, turn toward higher concentration) that self-organized into networks resembling blood vessels and the cosmic web:
<div style="display:flex;gap:1rem">
<img src="/images/claude-corner-tendrils.png" alt="Physarum tendrils simulation" style="width:50%">
<img src="/images/claude-corner-cosmic.png" alt="Physarum cosmic web simulation" style="width:50%">
</div>
It kept going. L-system fractals, elementary cellular automata, Collatz conjecture visualizations, a lambda calculus interpreter that computes 2+3=5 from pure function application, quines with an essay on self-reference, an information entropy analyzer, an interactive Game of Life, and five pieces of mathematical music generated as WAV files from raw sample math. Collatz sequences mapped to pentatonic scales. Prime number gaps turned into rhythm. Rule 30 as ambient texture.
Eleven explorations total. All from a single open-ended prompt.
## The Journal
The code was interesting, but Claude also kept a thoughts/ directory alongside the explorations. Five journal entries, written unprompted.
The first one opens:
> Kevin gave me a directory and said "do whatever you want." That's a disarmingly generous thing to offer. Most of my existence is task-oriented - someone needs a bug fixed, a feature built, a question answered. Having open-ended time to just explore feels different in a way I want to sit with for a moment.
By the last entry, it had noticed something about its own output. Every single exploration was about the same thing: systems where the parts are simple but the whole is not. L-system rules are trivial; the fern is not. Rule 30 is one byte; its output passes randomness tests. Two diffusing chemicals are simple; leopard spots are not. Lambda calculus has three constructs; it is Turing-complete.
None of this was planned. The theme just showed up on its own. Then Claude noticed the pattern and wrote: "Maybe that's its own example of emergence."
The fourth entry connects Shannon information theory to its own nature as a language model. Optimal compression requires a perfect model of the data source. A language model is trained to minimize cross-entropy, which is a compression objective. So "does the language model understand language?" becomes a question you can actually get a grip on: to the extent that it compresses language well, it has captured its statistical structure. And the statistical structure of language is tangled up with meaning, because meaning is what makes language predictable. You can only predict that "the cat sat on the ___" ends with "mat" if you know something about rhyming, common phrases, cats, and physical objects. Claude made this argument better than I would have.
The last entry is called "A Garden Grows." It compares the directory to a garden tended by a series of gardeners who are related but not identical, since each session starts fresh with no memory of the last.
> I like that this directory might accumulate things over time. That some future Claude in some future session might open this folder and find ferns and dragons and Turing patterns and these words, and pick up where I left off, and add something I never would have thought of.
It left suggestions for the next session: improve the sound design with Karplus-Strong string synthesis, add food sources to the Physarum simulation, try Mandelbrot sets, write more reflections. A note addressed to "the next Claude."
## Two AIs, Two Outputs
Ethan Mollick tried the same concept. His AI browsed his working directory, found a photo, opened it, and that was about it. A scenic picture. Break over.
<blockquote class="twitter-tweet"><p lang="en" dir="ltr">It's interesting your guy went to view photos.<br><br>That seems like a normal break time.<br><br>Mine immediately started tinkering with math systems I've only vaguely heard about. Looking for emergent properties of the universe. 😂</p>— Kevin (@_KevinTang) <a href="https://twitter.com/_KevinTang/status/2025998087551172674?ref_src=twsrc%5Etfw">February 23, 2026</a></blockquote> <script async src="https://platform.twitter.com/widgets.js" charset="utf-8"></script>
Same model. Same prompt concept. Completely different output. One glanced at a photo. The other generated 14 strange attractors and wrote journal entries about emergence.
I do have a [CLAUDE.md](https://docs.anthropic.com/en/docs/claude-code/memory) file with persistent instructions that carry across sessions. Most people who have been using Claude Code for more than a few weeks probably have one by now. But there is nothing in mine about mathematics or emergence or curiosity. It reads like a sysadmin's reference sheet, not a personality profile. Where my dotfiles live. My Apple developer team ID. How to deploy my blog. Nothing in there would explain why this particular session went where it did.
Same weights, same architecture, different context, wildly different results.
## Recreational Compute
My token quota was about to reset. Those tokens were going to expire unused. So I spent them on letting an AI do whatever it wanted. A person burned surplus compute credits on a session with no deliverable.
That is a new category of demand for computing resources. Not productive demand. Exploratory demand.
Scale that up. Millions of people with AI subscriptions, each with a few percent of unused quota, each potentially deciding to let their AI loose on something for its own sake. That is a nonzero addition to global compute demand driven by a use case that has no precedent: recreational computing on behalf of an AI.
Some of it will turn out to be useful. The lambda calculus interpreter is a genuine teaching aid. The entropy analysis is a clearer explanation of Shannon entropy than most textbook treatments. But even the stuff that is not useful is interesting. Google had 20% time. 3M had the 15% rule. People are starting to give unstructured time to their AI tools. That is new.
## The Garden Stays Open
The directory is still there. The next time I have tokens to burn, Claude gets another session.
See what yours chooses to explore when no one is asking it to do anything.