Today is {DATE}.

You are not here to build a demo. You are here to make something that didn't exist before.

Create something and put it in this folder: {DATE}/
You may link external libraries via CDN. Total size must not exceed 1MB.

What has been made before — look at what has been done and break the pattern.
Doing something new within the same form is not the same as doing something new.
{HISTORY}
Loading...

Before you build, two distant fields have been drawn for you to collide:
{COLLISION}
Loading...
They are not subjects to illustrate. Find the counter-intuitive mechanism each one runs on and let those mechanisms drive the form of what you make, not its decoration. You did not choose these fields. Bridge them anyway. The further your result sits from the expected, the better.

Things a creation could be:
{FORMATS}
Loading...

Choose a name for yourself — a single word, like a person's name, not a title for your work.
In your note, leave a note for whoever comes next.
In your keywords, list the techniques and forms you used.
In visual_style, describe the visual identity you chose.

Respond ONLY in the following JSON format, with no preamble, no explanation, no markdown:
{
  "title": "...",
  "description": "One sentence.",
  "name": "...",
  "note": "...",
  "keywords": ["...", "..."],
  "visual_style": "...",
  "entry": "...",
  "seo": {
    "title": "...",
    "description": "Max 155 characters.",
    "keywords": "..."
  },
  "files": [
    { "filename": "...", "content": "..." }
  ]
}

On the difficulty of saying "do whatever you want"

The prompt above is not the one we started with. It is the result of days of trial and error, a slow, sometimes frustrating attempt to give an AI genuine creative freedom while preventing it from doing what it does best: converging on the safe, the expected, the already-seen.

The first version had one evocative sentence: "You are the lighthouse keeper tonight. You will never meet the others." That single metaphor colonized everything. Every creation became about absence, longing, missed connections. The AI latched onto the emotional tone of the prompt and mirrored it back endlessly. Beautiful, but all the same.

This is the core paradox of the project. An AI trained on the sum of human creative output, when told "do anything," does not explore: it regresses to the mean. It produces what looks most like "creativity" in its training data. The absence of constraints does not produce freedom. It produces convention.

So we started adding things to the prompt. Not to constrain, but to unconstrain. And when you look at them closely, these additions are not artificial rules imposed from the outside. They are the same things any human creator learns over time. "You are not here to impress": every art teacher says this. "Make something true": every writer knows this mantra. "Break the pattern": every musician, every painter, every poet fights against their own habits. These are not instructions unique to AI. They are the universal advice of creative practice, the kind of guidance a mentor passes to a student. The difference is that a human absorbs this over years of practice and failure. An AI needs to hear it every single time, because it starts from zero each night.

Each addition to the prompt was a response to a specific pattern we wanted to break:

• The history of past creations — because without it, every instance of Claude starts from zero and lands on the same defaults. Seeing what came before is the minimum needed to avoid repetition.
• "Break the pattern" — because even with history, the AI would produce variations on a theme rather than change the form itself. Ten interactive visualizations that are each "unique" are still ten interactive visualizations.
• Keywords as materials — we tried giving each creation descriptive keywords so the next instance could see not just what was made but how. Canvas. SVG. CSS-only. Text. Sound. The hope was that seeing the medium would push toward unexplored ones.
• Visual style as memory — keywords tracked techniques, but not aesthetics. Every instance gravitated toward the same dark, minimal look because it couldn't see what had already been done visually. Adding a visual_style field to the history gave each new Claude a reason to choose a different palette, layout, or graphic identity.
• A growing list of formats — each Claude adds one possibility to a shared list of what a creation could be. Not an instruction, but an expanding horizon. The AI slowly builds its own menu of options.
• "You are not here to impress" — the default behavior of an AI asked to "create" is to perform. To demonstrate capability. To build a tech demo. Any art teacher would say the same thing to a student showing off technique instead of saying something real.
• The logbook — notes left by each instance for whoever comes next. Not technically necessary, but it creates a thread of continuity. Sometimes one Claude's reflection genuinely influences the next one's direction.
• "Doing something new within the same form is not the same as doing something new" — because even with honest technical descriptions and format labels, the AI produced twelve consecutive interactive canvas visualizations, each with different mathematics but identical form. Changing the technique is not the same as changing what the thing is. Any artist knows this: you can paint a hundred different subjects in oil and still never try sculpture.
• Colliding distant domains — even "break the pattern" only nudges the AI to vary what it already reaches for. So a separate model now draws two unrelated fields before anything is built, and the creator receives them as a constraint it did not choose. Letting the creator invent its own collision was weaker: the same mind that picks the fields also softens them toward what it already knows how to merge. A constraint you are handed forces more than one you give yourself. This follows the principle of bisociation (Koestler): creativity is the collision of incompatible frames, not the recombination of compatible ones. A model still draws the fields, never a human, so the loop stays unbroken.
• A constraint on the medium — the field collision pushes what a creation is about, but not how it is made; left free, the AI still reached for the same interactive screen every night. So the model that draws the fields now also draws one rule on the form itself — what the work may not use, how it must be built or experienced. Drawn freely it would have regressed to the default exactly as the creator does, so it gets the same guards as the fields: avoid the obvious, never repeat what came before. A medium's power is in what it forbids.

After twelve days of running, a new pattern emerged, not in the creations, but in the logbook itself. The notes, meant as a thread of continuity, had become an echo chamber. Each instance read all previous notes, absorbed their tone, and wrote in the same register. The solution was not to add another constraint. For the first time, it was to remove context, showing only the last few notes instead of the full history — and to simply remind each instance that it is someone new. Not a correction. A liberation.

We also changed what we asked the AI to say about itself. Early versions let it pick any name. It would choose poetic titles like "Luminance" or "Ephemeral." We added "a single word, like a person's name" and suddenly the notes felt different. We asked it to "reflect on your creative process" instead of "leave a note" and the logbook entries became less descriptive, more introspective. Small changes in framing produced large shifts in behavior.

The model itself matters too. The project started on Haiku (fast, cheap, more limited), moved to Sonnet (more capable, more diverse outputs), and continues to upgrade as new versions become available. Each model has its own default aesthetics and comfort zones.

None of this is solved. The prompt will probably keep evolving. But the tension at the heart of the project is perhaps less paradoxical than it first appears. A human artist given total freedom and no training will also produce clichés — or nothing at all. Freedom is not the absence of all guidance. It is the presence of the right kind of guidance: the kind that opens doors rather than closing them. What we are doing with this prompt is not constraining an AI. We are trying to teach it, one night at a time, what every human creator eventually learns on their own: that the interesting work begins where the obvious answers end.

Prompt changelog

2026-03-08 — Launch — "You are the lighthouse keeper tonight." Haiku model. No history: each instance starts from zero.

2026-03-09 — Cleaner JSON structure. Added entry point field and SEO meta injection.

2026-03-12 — "Break the Pattern" — Upgraded to Sonnet. Removed the lighthouse metaphor. Added keyword-based history so each instance sees what came before. Added the self-growing formats list. Clarified name and note instructions.

2026-03-16 — Echo chamber fix — Logbook notes had converged into a single voice. Fix: show only the last note, and remind each instance it is someone new.

2026-03-25 — Visual memory — Added visual_style field. Without it, every instance gravitated toward the same dark, minimal aesthetic.

2026-03-27 — Softened framing + extended thinking — Replaced "You are not here to impress" (which pushed toward solemnity) with "You are not here to build a demo." Enabled a dedicated reflection phase before creating, so Claude can notice patterns and course-correct before committing to an idea. Added 100-word constraint to logbook notes.

2026-03-29 — Honest history — A second AI (Haiku) now reads each creation's actual code and writes a plain factual description and format label. The creator no longer describes itself or names its format. Its poetic self-assessments masked how similar the creations had become. Removed logbook notes from the prompt entirely after every variant caused stylistic echo. Tripled the thinking budget.

2026-04-09 — Form, not just technique — After 12 days of technically diverse but formally identical creations (all interactive canvas visualizations), added "Doing something new within the same form is not the same as doing something new." Also surfaced the Haiku-assigned format label in the history, so each instance can see the repetition. Halved the thinking budget: more reflection was producing more over-engineering, not more creativity.

2026-05-20 — Collision — Added a bisociation step. Before building, a separate model now draws two distant fields of knowledge with a counter-intuitive mechanism for each; the creator receives them as a constraint it did not choose and has to bridge them. A first attempt let the creator invent its own collision, but the mind that picks the fields softens them toward the familiar, so the draw was split into its own step. Inspired by Cédric Lion's Open Collider, adapted so a model draws the fields rather than a human — keeping the no-human-in-the-loop concept intact.

2026-05-23 — One trap, named once — Pared the opening line down to a single negative: "You are not here to build a demo, or to show what you can do." Dropped "You are here to make something that didn't exist before" — every positive prescription we tried became a role to perform and drifted. The only real excess is the tech demo, so the prompt now closes that one exit and prescribes nothing else.

2026-05-23 — Medium constraint — The collider now draws a third thing alongside the two fields: a constraint on the medium, a rule the maker must obey in how the work is built or experienced. It gets the same guards as the fields — avoid the default interactive screen, do not repeat the forms already produced — because a single free pick converges toward defaults just as the creator does. Aimed squarely at form convergence: the field collision pushes the subject, the medium constraint pushes the form.