AI-built oil painting engine: 2,300 commits in 8 days as Claude Code runs a 'studio' of sessions
FotoRe_store · reddit · 2026-11-21
A developer open-sourced aibrush (AGPL-3.0), an oil-painting engine built for Claude and other strong models, with nearly all code written by Claude Code. The model lays strokes one by one, inspects, measures, and repaints — reproducing Manet's A Bar at the Folies-Bergère and the Mona Lisa.
Engine highlights
- Each stroke simulated per brush hair: pressure, depleting paint load, wet-on-wet dragging, spectral pigment mixing, lit impasto relief;
- The model gets "eyes": crops vs reference at any scale, squint view, tone/edge maps, numeric diffs, per-stroke reports;
- Trial strokes without committing, parallel variants, snapshot rollback; every painting is a replayable JSONL stroke journal; hard rule — every pixel comes from hairs/paint/pressure, no pasted or generated pixels.
How Claude built it
- Organized like a small studio: Engine, Acceptance, Editor, Tech, Coordinator sessions;
- Acceptance sessions never saw the author's reasoning; they run the full test suite and compare crops against real paintings — "look at the picture, not the number";
- Handoff happens via git documents (task file + ≤4KB state file per department), never chat; each session handles one item and closes around 200k context;
- 8 days: 2,300 commits, 19k lines of Python, 90 test files, 200 engine and 290 acceptance sessions.
Key finding: same engine, same model — one run produced a good painting with 9,777 strokes for 13M tokens, another a bad one with 970 strokes for 40M tokens. The README decides how the model paints. Documentation quality directly drives agent performance.
More from coding & agent
- sprite-stacking: open-source agent skill turns 4-view pixel art into spin-able fake-3D sprites — Jasonio · 2026-11-21
- Cargo adds GitHub integration: GTM agents open PRs, humans review before anything ships — ThePeterMick · 2026-11-21
- Taking AI Seriously: Borretti traces how programmer AI advice keeps retreating — zetalyrae · 2026-11-21
- Reverse-engineering Apple Watch: a single Claude Code prompt builds a physics watch face in 30 minutes — mtizard · 2026-11-21
- Agent RAG tweak: skip lossy summaries, extract answers, fetch 10 pages — 75% on validation — colinmcnamara · 2026-11-21
- us-legal-mcp: an MCP server exposing comprehensive US legislation to AI agents — modelcontextprotocol · 2026-11-21