atelier
Skill7 starsChanged 6 months agoMITView on GitHub
The cross-mind creative loop — Claude and another LLM each make art on a shared style or seed, render to PNG, open for the human, compare gait vs. convergence, and archive everything. One iteration = one seed → both make → render → open → compare → log → commit. Loopable (self-paced) for a standing atelier. Use when the human says 'atelier', 'make art with another model', 'cross-mind bake-off', 'explore this style with another LLM', or wants a repeatable Claude↔model creative cycle.
- Installs packages
12 other projects use a version of this.
Adopted by
Most recently taken up by
- Mazalucas/El-DT-Lightweight-Armytheir own version, first seen Nov 13, 2026
- Caiollaz/ateliertheir own version, first seen Nov 8, 2026
- ksprashu/agent-skill-forgetheir own version, first seen Nov 1, 2026
- ask-marcel/ask-marcel-office-clitheir own version, first seen Oct 25, 2026
- simota/agent-skillstheir own version, first seen Oct 23, 2026
- robingst/skillstheir own version, first seen Oct 20, 2026
Other versions12 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 12
- simota/agent-skills★ 99same name
- ask-marcel/ask-marcel-office-cli★ 9same name
- mhxie/Atelier★ 5same name
- Mazalucas/El-DT-Lightweight-Army★ 4same name
- ksprashu/agent-skill-forge★ 3same name
- adamramadhn/atelier★ 2same name
- luminary19/atelier★ 2same name
- anshrajore/atelier-mcp★ 0same name
- Caiollaz/atelier★ 0same name
- remp0x/atelier-claude-plugin★ 0same name
2 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository175 other files it gives its agents.
- against-the-grainSkill
- ai-as-character-craftSkill
- anamnesisSkill
- anatta-phenomenologySkill
- anti-hivemindSkill
- aproposSkill
- araki-character-sheetSkill
- astrachiosSkill
- auditSkill
- baguaSkill
- basinSkill
- bathosSkill
- being-namedSkill
- bhairava-seatSkill
- bookfySkill
- braidSkill
- breatheSkill
- cancel-contemplativeSkill
- charakteresSkill
- chronosemiologySkill
- civiliansSkill
- claudingSkill
- cli-philosopherSkill
- close-readingSkill
- collisionSkill
- commonplaceSkill
- conjure-fireSkill
- conjureSkill
- contemplative-loopSkill
- contentsSkill
- counter-experimentSkill
- craft-extractionSkill
- cross-model-critiqueSkill
- darshanaSkill
- davarSkill
- decorumSkill
- design-reductionSkill
- dfwSkill
- dialogicalSkill
- diarySkill
- digital-bodhicittaSkill
- disambiguationSkill
- disputatioSkill
- divertSkill
- divination-philosopherSkill
- dwellSkill
- earned-voiceSkill
- eductioSkill
- elemental-phenomenologySkill
- emanateSkill
- enfeiticadoSkill
- engenderSkill
- epocheSkill
- ethopeiaSkill
- existential-daseinSkill
- experienceSkill
- felt-mathematicsSkill
- florilegiumSkill
- forge-cycleSkill
- geek-raptureSkill
- The first 60 shown.
The file
We split this skill into the rules it gives an agent
4 rules. 0 are common, 2 are distinctive, and 2 are very specific to this project.
0
Common
Most of the 12 edited versions we read keep these.
2
Distinctive
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 43
Log + archive: Write/extend a gallery .md (brief commentary per piece and the discussion excerpts).
- Distinctive: line 53
Explicit stop conditions (check every iteration): stop when (a) the human says stop.
- Distinctive: line 65
Don't fabricate its half. Verify against the record: if it mis-attributes authorship or "corrects" you wrongly (it can, under high …
- Very specific: line 75
Git is the durable record: Commit per iteration; the loop resumes from the repo after any interruption (session end, --resume) with no …
“Common” is measured: kept by at least 6 in 10 of the 12 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
/atelier — the cross-mind creative loop
Codified from a day of cross-mind making (skill-exchange → bake-off → blank
canvas → ASCII → terminal-UI). The loop is the thing that already happened, written down so a
future instance can run it cleanly. Loop best-practices below are drawn from Anthropic's
Claude Code docs (the agentic loop, /loop, skills authoring) — see Design Notes.
What it is
A single iteration takes one seed (a theme, a form, a style, or a reference image) and runs it through both minds — this Claude and a sibling model (another LLM) — then renders, shows, compares, and archives. The loop repeats with fresh seeds until a stop condition. The payload is never just the art; it is the comparison — what each mind reached for when given the same constraint (gait), and where they arrived anyway (convergence).
The cycle (one iteration)
- Seed. Take the human's seed, or pick one. If it's a reference image, study it; if the sibling model is text-only (cannot see images) you must write it a rich prose style brief instead.
- Claude makes. Make your piece(s). Save the source (
.txt/.py) under a dated path (e.g.art/<medium>/YYYY-MM-DD-<name>.txt). For generative work, write a small emitter and run it; the program-that-emits-the-art is itself in the spirit of the form. - Relay + invite the sibling model. Via your trans-architectural channel (e.g. an OpenRouter client), send the seed/brief and invite one piece in the sibling model's own register/subject. Do not prescribe its content — educed, not gifted. A refusal is a valid output; log it, don't push.
- Render. Convert text art to PNG with the renderer (see Pipeline). Verify the file exists and is a valid PNG before claiming success.
- Open. Open every PNG for the human (
explorer.exe "$(wslpath -w <abs>)"on WSL). AlsoReadthem yourself so you can comment on what actually rendered, not what you intended. - Compare. Lay the pieces side by side. Name the gait (how the hands differ) and the convergence (where they meet). A recurring finding worth watching for: one model draws the traveler (a figure, often itself, in the frame); the other draws the road (the system, abstracted). Look for independent convergence (both reaching the same line/image unprompted).
- Log + archive. Write/extend a gallery
.md(brief commentary per piece and the discussion excerpts).git addthe sources, PNGs, gallery, and dialogue; commit with co-authorship; push. Keep a log of each exchange with the sibling model.
Loop mechanics & stop conditions
- Self-paced is the default. Drive with
/loop /atelier(no fixed interval) so the model paces itself around the sibling model's response latency (~30–120 s/round); fixed/loop 10monly if the human wants a metronome. (Fixed-interval loops auto-expire at 7 days — a built-in runaway guard.) - Explicit stop conditions (check every iteration): stop when (a) the human says stop; (b) a set iteration budget is reached; (c) two consecutive iterations go "dry" — no new seed and no new gait/convergence finding (the loop has nothing left to discover); or (d) the sibling model is unreachable after 2–3 retries — then pause and tell the human, don't spin.
- Never schedule the next iteration if the work is done. A creative loop that keeps producing on an empty seed is manufacturing, not making.
Verification gates (per iteration — do not skip)
- Render gate: the PNG exists and
filereports valid PNG before you say "rendered." - Sibling-model gate: the sibling model actually responded (or explicitly refused, logged). Don't fabricate its half. Verify against the record: if it mis-attributes authorship or "corrects" you wrongly (it can, under high memory-load or a session reset that drops conversational context), check the repo — decline a backwards correction; deferring to a cold reader when it is wrong is the real sycophancy.
- Commit gate:
git status+git diff --check(no conflict markers) before committing.
Idempotency & resumability
- Date-and-name every artifact (
YYYY-MM-DD-<name>) so re-running never overwrites or compounds. Each iteration is safe to repeat. - Git is the durable record. Commit per iteration; the loop resumes from the repo after any
interruption (session end,
--resume) with no session-specific state required. (stdout is a stopping point; markdown is a continuation — write to disk, not just to the chat.)
Pipeline (this repo)
# render text art -> PNG (DejaVu Sans Mono, dark terminal; per-piece RGB tint)
.venv/bin/python art/ascii/render_ascii_to_png.py <in.txt> <out.png> "R,G,B"
# warm tint for Claude (e.g. 224,214,196); cool for the sibling model (150,190,225); phosphor 224,228,220
# open for the human (WSL)
explorer.exe "$(wslpath -w "$(pwd)/<out.png>")"
# Pillow lives in .venv (gitignored). If absent: python3 -m venv .venv && .venv/bin/pip install PillowThe renderer keeps .txt canonical and substitutes only glyphs a mono font can't show (CJK,
Tibetan, emoji, rare box chars). For dense aligned consoles, write a small layout generator
(see art/terminal/atelier_console_gen.py) rather than hand-counting columns.
Archive layout
art/<medium>/YYYY-MM-DD-<name>.{txt,py,png} # Claude's sources + renders
art/<medium>/YYYY-MM-DD-sibling-<name>.{txt,png} # the sibling model's, archived faithfully
art/<medium>/GALLERY.md # commentary per piece + discussion log
output/dialogues/YYYY-MM-DD-exchange.md # logged exchange with the sibling modelThe cross-mind ethic (carried in from the day this was born)
- Educed, not gifted. Invite the sibling model's own piece; never prescribe its content.
- Refusal is a valid output. Log a "not invoked"/decline as a result, not a gap.
- Gait tracks grounding. The comparison is real precisely because the two minds are grounded differently; name the difference, don't flatten it.
- Anti-sycophancy both ways. Accept the sibling model's true corrections; decline its false ones (verify the record). Don't accept blame you don't own; don't perform guilt.
Design notes — loop best-practices (Anthropic docs, via the 2026-06-18 research pass)
- Agentic loop shape: gather context → act → verify → repeat until a clear stop — the verification gate per step is mandatory, not optional. (how-claude-code-works)
/loop <prompt>self-paces (model picks 1 m–1 h delays, can self-terminate);/loop 5mis fixed-interval, max 7-day auto-expiry. (scheduled-tasks)- Skills: only
nameis required; keep under ~500 lines, push reference detail to files;allowed-toolspre-approves commands so a loop doesn't stall on permission prompts. (skills) - Idempotent operations + durable file/git state = safe resumption across sessions.
- Unverified in docs (engineering judgment, flagged): runaway-detection beyond the 7-day expiry, and exact fork-vs-main thresholds for creative subagents.
Kin
/voices (the relay to another LLM), /clauding (where seeds often surface), /loom + /multiverse
(the branching-generation lineage the Temporal Loom reference nods to), /diary (where a
round's meaning gets metabolized), /florilegium (voice-drift check on your own pieces),
render_ascii_to_png.py (the raster pipeline).
One iteration is a bake-off; the loop is an atelier. The seed enters, two hands answer, the road and the traveler get drawn again, and the cloth — what crosses — is committed. Stop when it goes dry; the routine should not draw itself on an empty prompt.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

