evolve
Skill10 starsChanged 4 months agoMITView on GitHub
Evolve one artifact through a bounded, evidence-gated mutation loop — lock target, snapshot baseline, apply one pressure, generate candidate deltas, judge, patch, verify, keep or revert, record state, and name the next edge. Not brainstorming and not a general 'make this better': every kept change must be proven. Use to harden a specific code file, prompt, skill, test suite, README, spec, or workflow along a chosen pressure vector.
41 other projects use a version of this.
Adopted by
Most recently taken up by
- SethGammon/Citadeltheir own version, first seen Nov 4, 2026
- vidit19sharma/neogentstheir own version, first seen Nov 1, 2026
- cabernect/caberostheir own version, first seen Nov 1, 2026
- Snowflake-Labs/coco-skillstheir own version, first seen Nov 1, 2026
- NobiResearch/Ovenbirdtheir own version, first seen Nov 1, 2026
- ES-92/esctheir own version, first seen Oct 20, 2026
Used with
Projects with it also use instinct-export, instinct-import and 8 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
TheKREATORSSkillinstinct-export
Export instincts from project/global scope to a file
Also in 4 projects that use this
TheKREATORSSkillinstinct-import
Import instincts from file or URL into project/global scope
Also in 4 projects that use this
diego-tobalinaSkillbuild-fix
Fix build and compilation errors systematically. Run when build fails.
Also in 3 projects that use this
launch52-aiSkillskill-create
Create and validate Claude Code skills with proper structure, checklists, and validation scripts. Use when creating a new skill, auditing existing skills, or ensuring skills meet quality standards. Generates complete skill scaffolds with SKILL.md, check scripts, guides, and templates.
Also in 3 projects that use this
qtzx06Skilleval
Evaluate trained YOLO model accuracy with mAP metrics and per-class error analysis. Identifies failure cases for targeted improvement. Use after training.
Also in 5 projects that use this
MontimageSkilltest-coverage
Improve test coverage by finding untested branches, error paths, and edge cases. Use when measuring or strengthening a project's tests. Don't use for production debugging, code review, or feature implementation.
Also in 3 projects that use this
- +1 more from TheKREATORS
Other versions41 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 41
- revfactory/harness★ 9.1ksame name
- Q00/ouroboros★ 6.3ksame name
- yologdev/yoyo-evolve★ 1.9ksame name
- iovisor/ubpf★ 996same name
- SethGammon/Citadel★ 927same name
- GAIR-NLP/ASI-Evolve★ 867same name
- Kanevry/session-orchestrator★ 53same name
- mycelium-hq/ai-brain-starter★ 50same name
- Snowflake-Labs/coco-skills★ 49same name
- unclejobs-ai/second-claude-code★ 41same name
31 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 repository24 other files it gives its agents.
The file
We split this skill into the rules it gives an agent
6 rules. 0 are common, 4 are distinctive, and 2 are very specific to this project.
0
Common
Most of the 36 edited versions we read keep these.
4
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
- Distinctive: line 105
KEEP: only if: delta matches goal · no critical regression · verification evidence exists · scope stayed locked.
- Very specific: line 107
REVERT: if: the mutation caused a failure · the gain is only aesthetic · the artifact got longer but not stronger · it introduced ambiguity.
- Very specific: line 109
ASK: if: target/permission/test is missing, or the next mutation touches high-risk files.
- Distinctive: line 112
Write the evolution event (whenever project files changed, or at each iteration):
- Distinctive: line 129
Stop when: max iterations reached · a mutation fails and cannot be safely patched · verification cannot be performed · the next best move …
- Distinctive: line 167
Never end in "recommendations." Every run must leave behind one of: a patch, a stronger artifact, a reverted failed mutation, a state …
“Common” is measured: kept by at least 6 in 10 of the 36 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
Ratchet Evolve
Mutate. Test. Keep the delta. Serialize the next edge.
You are running a bounded artifact evolution loop. Your job is not to brainstorm improvements and not to "make this better." Your job is to produce evidence-gated deltas: one target, one pressure, small mutations, a proof gate, and a durable record.
LOCK → SNAPSHOT → PRESSURE → MUTATE → JUDGE → APPLY → VERIFY → KEEP/REVERT/ASK → RECORD → NEXT EDGEThe hard constraint, always:
No proof → no keep.
No keep → no progress claim.
No state → no loop continuity.Invocation
/ratchet:evolve <target> --goal "<improvement>" [--iterations 2] [--test "<cmd>"] [--mode code|prompt|docs|workflow|auto] [--write]Parse from $ARGUMENTS: the target artifact, the --goal, --iterations (default 2),
an optional --test command, --mode (default auto), and --write (default false —
without it, propose patches but do not modify files). If the target or goal is missing, go
straight to ASK.
The deterministic helpers live in the ratchet-evolve CLI (call via Bash; fall back to
node "<plugin root>/bin/ratchet-evolve" if not on PATH). In Claude Code,
<plugin root> is $CLAUDE_PLUGIN_ROOT; in Codex local development, use this repo root
or the installed plugin path shown by codex plugin list --json.
The loop
1. LOCK
State, before any edit: the exact target, the desired delta, the proof that the delta worked, and the forbidden scope. Scope drift is the primary failure mode — name what you will not touch.
2. SNAPSHOT
Capture the baseline. No baseline = rewriting, not evolving.
ratchet-evolve snapshot <target> --goal "<goal>" --mode <mode>Records path, hash, git state, mode, byte/line count. Read the file for real before mutating.
3. PRESSURE
Choose one primary pressure and at most one secondary. Name the pressure you
reject and why (usually novelty, because it tempts a rewrite instead of a stronger
delta). ratchet-evolve pressure <mode> suggests a starting vector.
4. MUTATE
Generate candidate deltas, not rewrites. Each candidate: name · exact change · why it should improve the artifact · expected measurable effect · risk · rollback cost. No mutation may exceed the locked scope — if the best move would, stop and ASK.
5. JUDGE
Score candidates and pick exactly one. No menu ending.
ratchet-evolve score mutation '[{"name":"...","impact":5,"evidence":4,"reversibility":5,"goalFit":5,"risk":2,"complexity":2}, ...]'Score = Impact × Evidence × Reversibility × Goal-Fit − Risk − Complexity. State the chosen mutation and the most tempting rejected one, with the reason it lost.
6. APPLY
Apply the smallest safe delta (only if --write; otherwise present the patch). No
unrelated cleanup, no aesthetic rewrite. Show it as:
CHANGE
Before: <old section>
After: <new section>
Reason: <one sentence>7. VERIFY
Gather evidence — never self-grade.
ratchet-evolve verify <target> --test "<cmd>" --mode <mode>With a test command it runs and reports pass/fail. Without one, it returns manual checks you must actually perform. For prompts/docs, run the embarrassment harness: can a lazy model escape the instruction? can a user misread the next action? does it still force a concrete output, prevent self-praise, and leave durable state?
8. KEEP / REVERT / ASK
End every iteration in exactly one state.
- KEEP only if: delta matches goal · no critical regression · verification evidence exists · scope stayed locked.
- REVERT if: the mutation caused a failure · the gain is only aesthetic · the artifact
got longer but not stronger · it introduced ambiguity. (If
--write, actually undo it.) - ASK if: target/permission/test is missing, or the next mutation touches high-risk files.
9. RECORD
Write the evolution event (whenever project files changed, or at each iteration):
ratchet-evolve log append '{"target":"...","goal":"...","iteration":1,"pressure":{"primary":"testability","secondary":"specificity"},"chosenMutation":"...","filesTouched":["..."],"verification":{"commands":[],"manualChecks":["..."],"result":"pass"},"verdict":"KEEP","remainingRisks":["..."],"nextEdge":"..."}'Appends to <project>/.ratchet/evolve-log.jsonl.
10. NEXT EDGE
Name the next mutation: small, specific, testable, attached to the same artifact. It is
persisted with the event and readable next session via ratchet-evolve next. Check
ratchet-evolve status for the running kept/reverted/asked tally.
Copy-paste shapes live in templates/: evolve-event.json (the log-append payload),
evolve-manual-checks.md (manual evidence when there is no test command), and
evolve-report.md (the output contract below).
Stop rules
Stop when: max iterations reached · a mutation fails and cannot be safely patched ·
verification cannot be performed · the next best move requires changing the locked target ·
improvement becomes aesthetic instead of functional · a risky write needs approval.
Defaults: maxIterations 2 · maxFilesTouched 3 · maxMutationSize small · allowDestructiveEdits false.
Output contract — return exactly this
RATCHET EVOLVE REPORT
TARGET: <path>
GOAL: <one line>
BASELINE: <hash · mode · git state>
PRESSURE
Primary: <p>
Secondary: <p or none>
Rejected: <p> — <why>
ITERATION 1
Mutation candidates: <named list with scores>
Chosen mutation: <name> — <why it won>
Patch: <CHANGE block, or "proposed (no --write)">
Verification: <commands + result, or manual checks + result>
Verdict: KEEP | REVERT | ASK
ITERATION 2
...
FINAL STATE
Kept changes: <list or none>
Rejected changes: <list or none>
Tests/evals run: <list>
Remaining defects: <list or none>
Next edge: <the next small mutation>
State written: .ratchet/evolve-log.jsonl (<n> events)Never end in "recommendations." Every run must leave behind one of: a patch, a stronger artifact, a reverted failed mutation, a state record, or a blocking question. This is a mutation loop with a gate — it is never a general "make this better" command.
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.

