bake-off
Skill0 starsChanged 36 days agoMITView on GitHub
Design tournament for decisions that are expensive to reverse (chunking strategy, index schema, retrieval architecture). Independent fresh-context candidates, cross-critique, and — wherever the decision is measurable — the tuning eval set picks the winner. No model-as-judge scoreboards, ever.
2 other projects use a version of this.
Adopted by
Most recently taken up by
- CorvusWing/CorvusWingtheir own version, first seen Nov 8, 2026
- robdennis/claude-skillstheir own version, first seen Apr 15, 2026
Other versions2 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 2
- CorvusWing/CorvusWing★ 1same name
- robdennis/claude-skills★ 0same name
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 repository4 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- phase-gateSkill
- plan-gateSkill
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 1 is distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
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 10
Use this only for decisions that are genuinely expensive to reverse — a bake-off costs several agent runs and possibly an eval cycle.
- Very specific: line 29
Write the brief: Create docs/designs/NNN-bakeoff-<slug>.md (next NNN.
- Distinctive: line 67
Loser becomes the watchdog: When the winning design is implemented, hand the losing candidate's critique list (and the loser design itself) …
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
You are running a design bake-off for: $ARGUMENTS
Use this only for decisions that are genuinely expensive to reverse — a bake-off costs several agent runs and possibly an eval cycle. For anything cheaply reversible, decide normally and log it in docs/decisions.md.
Hard rules for every bake-off
- The judge is the eval set, not a model. A model grading model output at n=1, with no ground truth and no blinding, is a vibes ledger — this harness rejected that pattern deliberately (see docs/harness.md). If no metric can decide, the human decides; a model never does.
- Candidate selection uses the tuning/golden set ONLY (for this repo:
eval/golden_set.jsonl). The held-out set is never an input to selection — using it to pick a design burns it as a headline metric. - Candidates get the brief and nothing else. Independence of context is the active ingredient: a candidate that saw another candidate's reasoning, or the conversation that framed the problem, is contaminated.
Procedure
- Write the brief. Create
docs/designs/NNN-bakeoff-<slug>.md(next NNN; start fromdocs/designs/TEMPLATE.md). It contains ONLY: the problem statement, the constraints (from CLAUDE.md + relevant decisions.md entries, cited by number), the decision criteria, and the exact eval command + metric that will judge (or the explicit statement that the decision is unmeasurable and the human judges). Confirm the brief with the user before spawning anything — the brief is the whole contest. - Generate candidates independently. Spawn two design agents in
parallel (Agent tool, general-purpose), each receiving only the brief's
file path and an output path (
## Candidate A/## Candidate Bsections of the brief, or sibling files for long designs). Fresh context each — do not share this conversation. If a second vendor is available (codex exec --sandbox read-only), it may author one candidate from the same brief — pointers only, never pasted corpus text. Vendor diversity adds variety of failure modes; context independence is what actually matters. - Cross-critique. For each candidate, spawn a fresh plan-auditor pointed at that candidate and the brief's criteria. Append each failure-mode list under the candidate it attacks. Verify findings before trusting them — critics can be wrong.
- Judge.
- Measurable: implement each candidate minimally, isolated from the
working tree and from each other (git worktrees; separate persist
directories for anything indexed — never the real
chroma_db/). Run the brief's exact eval command per candidate. The metric table picks the winner. If generation-stage API calls are needed, state the cost and get the user's go-ahead first. - Not measurable: present both candidates plus critiques to the user with a recommendation and the reasoning. The user picks.
- Measurable: implement each candidate minimally, isolated from the
working tree and from each other (git worktrees; separate persist
directories for anything indexed — never the real
- Record. Winner, numbers (or the human's stated grounds), and the loser-as-rejected-alternative go into a docs/decisions.md entry, linking the brief. The losing candidate stays in the design doc — decisions.md demands "what was rejected", and the bake-off already wrote it down.
- Loser becomes the watchdog. When the winning design is implemented, hand the losing candidate's critique list (and the loser design itself) to the phase-gate pressure-tester as extra attack surface. That is the one part of the "losing model tears it down at checkpoints" idea worth keeping: the loser's design encodes a different set of assumptions, which is exactly where the winner's blind spots live.
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.

