evolve
Skill0 starsChanged 42 days agoMITView on GitHub
Improves code, an interface or a document in scored generations instead of blank-canvas rematches. Each round breeds candidates from the top two versions of the round before, gates every candidate on the project's build, type check, linter and tests, and for an interface also a clean render, then one blind judge in a separate agent scores the pool on a frozen rubric, the parents included unlabelled as anchors. Stops the first round the best score does not rise, never runs past 6 rounds, then applies the winner and reports a full analysis with a round-by-round score curve. Use when the user says "/evolve", "evolve this", "keep improving it round by round", "iterate until the score stops rising", "carry the best forward", "hill-climb this", or the Turkish "round round iyilestir", "en iyileri bir sonraki rounda aktar", "puan artmayana kadar devam et", "evrimlestir". For rebuilding a screen from a blank canvas each round use compose; for the interaction flow use humanize; for a fixed rule pass use lint.
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 10 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
AWSSkillmap
Map verified service capabilities against customer-provided control framework and threat model. Produces mapping-results.json, controls-catalog.md, and framework-mapping.md. Supports pluggable frameworks: CCMv4, NIST 800-53, CIS Benchmarks, ISO 27001, etc. Split into 4 focused sub-skills for better accuracy; this skill orchestrates them.
Also in 5 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
- +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.
The file
We split this skill into the rules it gives an agent
2 rules. 0 are common, 2 are distinctive, and 0 are very specific to this project.
0
Common
Most of the 36 edited versions we read keep these.
2
Distinctive
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 72
Write the leader's new score beside its old one.
- Distinctive: line 79
Not higher (lower, tied, red-lined, every candidate pulled back by the pre-crown check, all gated out) → stop.
“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
evolve — breed the best forward
Breed candidates from the best versions so far, gate them, let a blind judge score the whole pool, carry the top two forward. The first round that does not raise the best score ends the run.
Read ../_shared/tournament.md first. It defines the loop and binds this skill — gate, parallel
builders, one blind pooled judge, red lines, pre-crown check, apply and verify — except for these
overrides:
- §1 — R0 is also gated and scored alone by the judge as the curve's baseline (§1 below).
- §4.1-§4.2 ideas — candidates are bred from the leader (§2 below), not built from fresh ideas.
- §4.4 pool — both parents, leader and runner-up, sit in the pool as anchors, not one champion.
- §4.6-§4.8 throne — ranks 1 and 2 carry forward; the best candidate must strictly beat the best parent of the same pass (§3-§4 below).
- §5 cap — 6 rounds; the end sentence reads no increase at round N or round 6 reached.
0. Target and rubric
The argument is the target (/evolve the pathfinding cache); without one, ask one question: which
file, function, screen or document. Pick the rubric by what is improved and freeze it with the
target (tournament.md §1):
| Target | Rubric | Gate: tournament.md §3, plus | Pre-crown check (tournament.md §4.7) |
| Code | tournament.md §2 | — | — |
| How an interface looks and reads | ../compose/SKILL.md §3, its red lines in §5 | ../compose/SKILL.md §2(c) | the alignment audit (../compose/SKILL.md §4) and ../compose/references/checklist.md, with the Layout cap of compose §4 |
| How an interface behaves in a person's hands | ../humanize/SKILL.md §4, its red lines in §5 | ../humanize/SKILL.md §2(c), then the walk (../humanize/SKILL.md §3) whose traces the judge reads | ../humanize/references/interaction.md, with the caps of humanize §4 |
| A document | tournament.md §2, tailored before round 1 | whatever renders or validates it | — |
For the interface rows, also freeze what compose / humanize freeze in their §0, render at the
round's viewports (tournament.md §4.3) and load the design skill they name. Their red lines are
written against "the champion"; here read the best parent in the same pass for it. Every parent,
leader or runner-up, went through the pre-crown check before it was carried and stays capped by its
findings.
1. Round 0
Work folder <scratchpad>/evolve/<target-slug>/. R0 is the current version in r0/, gated and
scored alone by the blind judge (§3) as the curve's baseline. If nothing exists yet, there is no R0: round 1 builds
three fresh candidates from the frozen target, as tournament.md §1.1 and §4.1-§4.2, judged with no
parent in the pool and no stop comparison. Down the ranking (§4.7 there), the first candidate to
survive its pre-crown check is crowned leader and the next survivor becomes runner-up; leader and
parent comparisons start at round 2. None survives → stop, nothing applied.
2. The round
Parents of round N: the leader (rank 1 of the previous pass) and the runner-up (rank 2). Round 1 has only R0 (or no parent at all, §1).
Breed three candidates in r<N>/c<k>/, all from the leader, each with one move written down
before any edit: two fixes aimed at the leader's weakest criteria, and one graft of the runner-up's
idea where it beat the leader (a third fix when there is none). No move already tried in an earlier
round. Builders run in parallel as in tournament.md §4.2, each handed the leader and its own move
instead of a fresh idea.
Gate each candidate (the table's gate column, tournament.md §3). All out → no increase (§4).
3. The judging pass
tournament.md §4.4-§4.6, with the surviving candidates plus both parents in the pool, no hint
which is a parent. Ties go to the parent. Before ranks 1 and 2 are fixed, each candidate in line for
either goes through the row's pre-crown check (§4.7 there), highest-ranked first; its caps re-rank it,
and the check moves down the pool until both places hold checked versions. Ranks 1 and 2 become the
next leader and runner-up, parent or candidate.
Write the leader's new score beside its old one; the gap is the judge's drift.
4. Stopping
Compare the best candidate with the best parent in the same pass:
- Strictly higher → the score rose; next round.
- Not higher (lower, tied, red-lined, every candidate pulled back by the pre-crown check, all gated out) → stop; the best parent is champion.
- Round 6 finished → stop, even if still rising.
Append each round to rounds.md before the next: round, moves, gate, the pass's ranking with
scores, pre-crown check, carried forward. Then apply and verify as tournament.md §5 steps 1-4; a
champion of R0 changes nothing, and say so.
5. Final analysis
tournament.md §6, with one row per round:
| Round | Moves | Gate | Pool (best → worst) | Best parent | Best candidate | Pre-crown check | Δ | Carried forward |
A parent's score is the one from that round's pass. How we did it traces the winner's lineage
back to R0 — the move each generation added, which graft survived. Possible mistakes names at
least judge drift, the local optimum of building on one line (no blank-canvas run — compose or
humanize does that), and whatever the gate could not check. The sentence after the table: no
increase at round N, or round 6 reached.
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.

