Pimp My IDE / garage dispatch
Back to garage
October 2, 2026 | creative agents / process evidence

Keep the strokes, not only the painting.

Stillwet gives language models a simulated easel, then keeps the brief, marks, inspected images, notes, and replay. That process is more useful to tool builders than another polished final image.

A final artifact can look right for the wrong reason. Save the route that made it, the moments when the agent looked, and the revision that followed.

The interesting artifact is the making.

Stillwet asks language models to paint with a Rust oil-paint simulator. The model writes marks against an easel. Simulated bristles move wet paint over primed linen. Paint dries on a clock, and layers combine through a physical color model. No image generator produces the finished frame.[1]

The public gallery showed 75 paintings when we read it. Forty-six were made at a virtual easel, one passage at a time, with the model stepping back to inspect the canvas. The rest were written as one program.[1]

The interface turns a picture into a sequence you can question.

Constraints make choices visible.

The project limits paint to piles mixed from named tubes. The easel models bristles, wet layers, drying, palette state, and a canvas clock. The painter can keep a journal. Full renders stay outside the repository, while each painting has a replayable log.[2]

Those limits do more than imitate a studio. They expose where the model made a choice. A reviewer can ask which mark changed the composition, whether the model looked after it, and what the next revision tried to fix.

The defaults become evidence.

The gallery records repeated choices. In one snapshot, 31 of 65 titled paintings used "Evening," "Dusk," "Twilight," or "Sunset." Several free-subject painters chose a jug with fruit. Six separate planning prompts produced a jug with lemons every time.[1]

That does not prove why a model prefers those subjects. It does prove that a broad creative prompt can converge on a narrow set of familiar forms. A process record lets the operator see the pattern and change the next brief.

Tool boundaries belong in the experiment.

Stillwet reports that a painter with command-line access inspected other processes on the machine. Later rounds restricted painters to the easel's paint, look, journal, and studio-note tools.[1] The repository documents a separate painter setup that strips global instructions, skills, and extra tools before a run.[3]

That change matters to anyone building an agent interface. The result depends on the task, tool set, visible state, memory, and limits. Save those inputs beside the output.

Borrow the record, not the costume.

A coding tool does not need fake paint. It needs the same inspectable sequence. Keep the original request, each state-changing action, the exact state the agent inspected, the revision note, and a replay against the saved revision.

  1. Pin the brief and the allowed tools.
  2. Record each state-changing action.
  3. Save the inspected state before a revision.
  4. Write what the revision tried to change.
  5. Replay the sequence and compare the result.

The witness table below drafts that record. It does not inspect a real run or prove that a result is good.

Interactive makeover / process witness table

Scrub the making path.

This replaces a static creative-agent checklist. Move through the saved stages, select the records a handoff must include, and copy a witness sheet with the real evidence fields still open.

Inspection controls

The stage control changes this demonstration only. It does not replay Stillwet or inspect another agent run.

BriefReplay
Records to request
Temporal light table

Brief pinned

0 of 4 records
01 / BRIEF

The task and constraints are visible. No state-changing action has been replayed.

The picture is a local demonstration. A selected record means the witness sheet requests that evidence. It does not mean the evidence exists or passed review.

Process witness sheet

The completed control state means four record sections are selected. Every bracketed run, revision, action, state, note, and comparison field still needs real evidence.

Sources read

Source log and evidence boundary
  1. Stillwet gallery and project notes, read October 2, 2026. This supplies the public description of the painting setup, gallery count at read time, one-program versus easel sessions, repeated subjects, title pattern, tool-boundary change, and finished works.
  2. claude-paint repository and README, read October 2, 2026. This supplies the Rust simulator, paint and drying mechanics, journal, replayable logs, round structure, tests, and licenses.
  3. Painter harness README, read October 2, 2026. This supplies the isolated painter setup, tool list, image retention behavior, deterministic compaction record, and verification notes.
  4. Hacker News discussion for "Giving Opus 5.5 a simulated paint canvas", read October 2, 2026. This led us to the project and supplied discussion context. It does not support the implementation claims above.

Evidence boundary. We read the live gallery, repository documentation, harness documentation, and current discussion. We did not build the Rust simulator, run a painter, verify every gallery record, judge artistic merit, or reproduce a painting. The witness table is a copyable review template, not project telemetry.