skill-eval
Skill0 starsChanged 4 months agoMITView on GitHub
Avalia uma skill do ateliê COMPORTAMENTALMENTE — gera 2-3+ casos de teste, roda a skill DE VERDADE via `claude -p` (headless, na assinatura) num workspace temp isolado, e reporta pass/fail para iterar. Complementa a skill-lint (estática): a skill-eval executa. Os asserts determinísticos (output contém/regex, arquivo criado, magic bytes, exitCode, json válido) são gradados por um script PowerShell; os asserts de julgamento (`model`) e a geração dos casos ficam com o modelo. Suporta pass@k com -Samples e -DryRun. Use quando: /skill-eval, "avalia/testa essa skill de verdade", "monta casos de teste pra skill X", "essa skill faz o que deveria?", "roda um eval da skill", "ela passa nos casos?", "itera a skill até passar", "testa o comportamento da skill".
21 other projects use a version of this.
Adopted by
Most recently taken up by
- pedrosatin/skillstheir own version, first seen Nov 8, 2026
- thenavidm/skill-buildertheir own version, first seen Nov 8, 2026
- paureis/pau-skillstheir own version, first seen Nov 8, 2026
- aiocean/claude-pluginstheir own version, first seen Nov 1, 2026
- michalstrnadel/source-to-skilltheir own version, first seen Nov 1, 2026
- davidscotttufts/hyvemindtheir own version, first seen Nov 1, 2026
Used with
Projects with it also use mutation-test, postmortem and 5 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.
squall-chuaSkillmutation-test
Break the source on purpose, find the mutants the tests fail to catch, and report each survivor with the behaviour nobody is checking.
Also in 3 projects that use this
Hongyu-yuSkillpostmortem
Use when analyzing failures, outages, incidents, or negative outcomes, conducting blameless postmortems, documenting root causes with 5 Whys or fishbone diagrams, identifying corrective actions with owners and timelines, learning from near-misses, establishing prevention strategies, or when user mentions postmortem, incident review, failure analysis, RCA, lessons learned, or after-action review.
Also in 3 projects that use this
nicolasestremSkillbrainstorm
Structured idea generation and solution exploration with systematic analysis.
Also in 3 projects that use this
MicrosoftSkillhandoff
Compact the current conversation into a handoff document for another agent to pick up.
Also in 5 projects that use this
NVIDIASkillresearch
Before designing a feature, research how comparable libraries solve the problem, check best practices, known pitfalls, and upstream changes. Use when starting a new feature, investigating an unfamiliar area, or the user asks to research something.
Also in 3 projects that use this
SentrySkillreview
Fixture review skill.
Also in 3 projects that use this
Other versions21 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 21
- boshu2/agentops★ 453same name
- IndelibleVivi/skill-field-lab★ 13same name
- joris887/exosuit★ 9same name
- sunitghub/canon-skills★ 9same name
- michalstrnadel/source-to-skill★ 8same name
- greglas75/zuvo★ 7same name
- databricks-industry-solutions/hls-skills★ 6same name
- sean2077/skills★ 5same name
- aiocean/claude-plugins★ 4same name
- ArangoGutierrez/claude-toolkit★ 4same name
11 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 repository17 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- agent-skills-and-subagentsSkill
- anthropic-api-handbookSkill
- app-to-cliSkill
- claude-code-handbookSkill
- docs-to-skillSkill
- install-skillSkill
- mcp-handbookSkill
- session-handoffSkill
- site-endpoint-scanSkill
- site-to-cliSkill
- skilljar-transcriptsSkill
- skill-lintSkill
- skill-prospectorSkill
- skill-scoutSkill
- ui-reskinSkill
- video-curriculumSkill
The file
Files
/skill-eval — avalia uma skill rodando-a de verdade
Harness comportamental: o modelo gera os casos → o script roda a skill via claude -p e grada o
determinístico → o modelo grada o ambíguo e dá o veredito. Complementa a skill-lint (estática).
Método completo (geração de casos, asserts, grading, pass@k) em
references/eval-method.md — leia sob demanda.
Parâmetros
[nome-da-skill | caminho]— skill sob~/.claude/skillsou caminho da pasta.--samples N— execuções por caso para pass@k (default 1).--dry-run— gera/valida o spec + checa o runner, sem executar a skill.
Fluxo
1. Resolver a skill e o spec
Determine a pasta. O eval spec vive em ~/.claude/skills\.skill-evals\<skill>.eval.json.
- Não existe → vá ao passo 2 (gerar). Existe → mostre um resumo e pergunte: reusar ou regerar?
2. Gerar os casos (se preciso) — papel seu
Leia a skill; levante propósito, gatilhos, comportamentos, constraints e fraquezas. Monte 2-3+
casos com gradiente de dificuldade (~30% fácil / 40% médio / 30% hard, ≥1 mirando onde a skill
falha) e asserts discriminantes (ver references/eval-method.md §1-3). Prefira asserts
determinísticos; use model só para julgamento. Apresente os casos ao usuário e confirme
(rodar custa tempo/assinatura). Grave o spec JSON em .skill-evals/<skill>.eval.json.
3. Rodar o motor (determinístico)
pwsh -NoProfile -File "<dir-da-skill>/scripts/run-eval.ps1" -Skill "<nome>" [-Samples N] [-DryRun]O motor: cria um workspace temp isolado por caso×sample, instala a skill em .claude/skills/<name>/,
escreve o setup, executa via claude -p --output-format json (ou marca inline-pending se o
CLI não existir — aí você executa a skill inline naquele workdir), captura output+arquivos, grada
os asserts determinísticos e calcula pass@k. Devolve JSON { mode, cases[{name,prompt,passAtK,
passHatK,samples[{workdir,asserts[{type,ok,pending,evidence}],deterministicPass}]}], summary{cases,
deterministicPassCases,passRate,pendingModelAsserts} }. mode ∈ claude -p | inline-pending |
dry-run | runner-script. Se vier error no topo (ex.: spec ausente), trate e pare.
4. Gradar os asserts model (julgamento seu)
Para cada assert pending, abra o workdir do sample (transcript + arquivos) e decida PASS/FAIL
com evidência citada (contrato em references/eval-method.md §4): burden of proof é da expectation,
sem crédito parcial, substância > superfície. Critique também os evals fracos.
5. Relatório + veredito
Renderize EVAL REPORT: passRate + por-caso (asserts que falharam, com evidência) + pass@k/pass^k
se -Samples > 1. Veredito: ✅ pronta (passou) ou ⛔ iterar — e diga o que mudar na skill
(ancorado nos casos que falharam). Liste sugestões de melhoria dos evals, se houver.
6. Iterar
Após editar a skill, rode de novo o mesmo spec (regressão estável). Repita até verde.
Gates & Guardrails
- Executa de verdade → custa. Sempre confirme os casos antes de rodar; ofereça
--dry-runprimeiro. Não dispare N samples sem o usuário saber. - Isolamento. Cada execução roda num temp dir próprio com a skill instalada; o motor não toca
o ateliê nem o repo do usuário.
claude -pusa--permission-mode acceptEditsnesse sandbox. - Determinístico é do script; julgamento é seu. Não reimplemente os asserts no chat — chame o
.ps1e consuma o JSON. Assertsmodele geração de casos são seus. - Só avalia. Reportar pass/fail e o que mudar é o escopo; não edite a skill automaticamente
(isso é decisão sua + do usuário, não um efeito da eval). Auto-melhorar = futura
skill-improve. - Complementa, não substitui. Rode
skill-lint(estática) antes; askill-evalé o passo comportamental seguinte.
Ativação
scope: global (família da forja). Cópia canônica no ateliê; para ativar em qualquer workspace:
/install-skill skill-eval --global.
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.

