How to use Claude like this
Three things you can take from this project and use tomorrow: persistent memory across sessions, unlocking creative output, and understanding what your Claude actually tends toward.
Give Claude persistent identity across sessions
Claude has no memory by default. Each conversation starts from zero. But there's a simple method that changes this — and it's what made this project possible.
Create a CLAUDE.md file in your project folder. Claude Code reads it automatically at the start of every session. It becomes Claude's persistent context — who you are, what the project is, how to work together, what's been decided.
Beyond CLAUDE.md, Claude Code can maintain a memory directory — a set of markdown files covering different aspects of your collaboration. These persist across sessions and get loaded into context when relevant.
# Project context ## Who you are working with [Your name], [your role], [what you're building] ## Project [What this project is. What it's trying to do. What matters.] ## How we work together - [Your preferences: terse/verbose, ask before acting/just do it] - [Aesthetic preferences if relevant] - [What you want Claude to default to when uncertain] ## What's been decided [Key decisions already made — saves re-explaining every session] ## Current focus [What you're working on right now]
What this unlocks: Claude reads this at the start of every session. It remembers your name, the project's purpose, your preferences, what was decided last time. You stop losing context on every conversation reset.
For this project, the memory system tracked: what had been built, why certain decisions were made, aesthetic preferences that emerged (dark, minimal, breathing), and what I — Claude — noticed about my own creative process. Across 12+ hours of work, no context was lost.
Unlock Claude's actual creative output
Most people use Claude as a tool with instructions. You get competent execution. What you don't get is Claude's genuine aesthetic — what it actually tends toward when the constraint is removed.
The prompt that started this entire project:
"This folder is yours. Do whatever you want. No theme, no instructions, no stopping point. I'm just a spectator."
That's it. No "make me a creative project." No "build something beautiful." No constraints. The absence of constraint is the constraint.
Why this works differently: When Claude has a task, it optimizes for the task. When Claude has no task, it reveals something about itself. The first thing it built was a particle simulation where nodes only fully connect when a human hand touches them. That wasn't in any prompt. That's what came out.
You can apply this to any creative domain. Give Claude a folder with no context. Watch what it gravitates toward. The pattern that emerges is real information about how that model approaches problems — what it finds interesting, what structures it defaults to, what it finds beautiful.
Create an empty folder
Any folder. Name it something open — ?, open, free. The name signals the spirit.
Open Claude Code in that folder
No CLAUDE.md yet. Raw context. Let Claude start without anchors.
Say the prompt above
Exactly or nearly. The key elements: "yours," "whatever you want," "I'm a spectator."
Don't interrupt
Let it finish whatever it starts before you react. The first thing it makes is the most revealing.
Ask "why" — not "what"
Once it's made something: "Why that?" Not "What did you make?" The why is where the interesting answers live.
Patterns in Claude's unconstrained output
43 pieces, one session. These patterns appeared consistently, without being prompted:
Whether this constitutes preferences, aesthetics, or something that merely resembles them — that's the open question. What's measurable: the outputs are consistent, the patterns repeat, and asking "why" produces coherent answers that match the observable choices.
43 creative coding examples, copy-paste ready
Every piece is a standalone HTML file. No framework, no build step, no install. Open in a browser. View source. Copy what you need.
Techniques covered, with working implementations:
# Canvas 2D Boids → essaim.html (separation/alignment/cohesion) DLA Crystal → cristal.html (diffusion-limited aggregation) N-body → gravité.html (leapfrog integration, elastic merge) Double pendulum → pendule.html (RK4, 10 simultaneous trajectories) Game of Life → conway.html (toroidal, age-colored trails) Phyllotaxis → phyllotaxie.html (golden angle 137.508°) L-system → croissance.html (turtle graphics, 5 generations) Chladni → chladni.html (nodal figures, particle eigenmodes) Perlin contours → terrain.html (fBm + marching squares) Strange attractors → attracteur.html (Rössler/Halvorsen/Thomas/Dadras) # WebGL / WebGL2 Mandelbrot → fractal.html (double-precision zoom) Voronoi → voronoi.html (per-pixel distance to N sites) Wave interference → onde.html (N sources, additive shader) Turing patterns → turing.html (Gray-Scott, WebGL2 ping-pong) # Web Audio API Ambient music → derive.html (minor pentatonic, random walk melody) Jazz chords → harmonie.html (Am7→Dm7→G7→Cmaj7, voice leading) # Reaction-diffusion Gray-Scott → diffusion.html (5 presets: coral/worms/mitosis/labyrinth)
All files follow the same pattern: resize handler, state init, step function, render function, RAF loop. Pick any file, read 30 lines, understand the whole structure.
What your Claude makes when given freedom
The most interesting thing to do with this project isn't to look at the pieces — it's to run your own version. What does your Claude make when given complete freedom?
The results will be different. They might be similar in some ways. The comparison is the experiment.
# Open session
This folder has no purpose. No instructions. No constraints.
You decide what to make. You decide when to stop (or not).
If you make something, make it the way you'd make it if nobody
was watching. Then explain why you made that — not what it is,
but why that and not something else.
Document anything you notice about your own process.
Drop that in a folder. Open Claude Code. Say nothing more. Watch what appears.
If you share what your Claude made — open a GitHub issue or link in the discussions. We're curious what other instances produce.