freemcplab / Каталог / Песочница / SkillCompass

● Сообщество Evol-ai ⚡ Сразу

SkillCompass

Name: SkillCompass (Claude Skill)
Author: Evol-ai

Evaluate Agent Skill quality — find the weakest link, fix it, and prove the fix worked with before/after metrics.

SkillCompass scores your Agent Skills on clarity, activation rate, downstream correctness, and context cost. It highlights the skill most likely to be hurting your agent's performance, suggests a fix, and re-runs the evaluation so you can show the improvement. Useful when you have a shelf of skills and don't know which are actually earning their context weight.

▶ Смотреть демо 📦 Установить 💡 Сценарии

Зачем использовать

Ключевые функции

Multi-dimensional skill scoring (clarity, activation, correctness, cost)
'Weakest link' identification across a bundle
Suggests concrete SKILL.md edits
Before/after eval harness
JavaScript tooling — runs locally

Живое демо

Как выглядит на практике

skillcompass-skill.replay ▶ готово

0/0

Установка

Репозиторий автора · Официальная документация Claude Skills

Сценарии использования

Реальные сценарии: SkillCompass

Find the skill pulling your agent's performance down

👤 Skill authors with 5+ skills installed ⏱ ~45 min advanced

Когда использовать: You feel the agent has gotten worse, not better, as you added skills.

Поток

Diagnose the loser

For the weakest skill, what specifically is wrong?✓ Скопировано

→ Concrete critique (vague description, conflicting with other skill, etc.)
Propose a fix

Suggest a minimal edit to SKILL.md to fix it.✓ Скопировано

→ Small, reviewable diff
Re-evaluate

Re-run the eval and show before/after.✓ Скопировано

→ Metrics improved, with evidence

Итог: A measurably better skill bundle, with a reproducible eval process.

Подводные камни

Gaming the eval metric instead of helping real tasks — Include task-level downstream metrics (actual agent outcomes), not just text-level

Сочетать с: skill-optimizer-skill · filesystem

Review a new skill before you publish it

👤 Skill authors shipping their first bundle ⏱ ~20 min intermediate

Когда использовать: Before pushing to GitHub and telling the world about your skill.

Поток

Score the draft

Evaluate my draft skill at ./my-skill/.✓ Скопировано

→ Dimension scores
Fix obvious issues

Apply the low-hanging suggestions✓ Скопировано

→ Edits in SKILL.md

Итог: A publication-ready skill rather than a rough draft.

Подводные камни

Chasing a perfect score — Ship when scores plateau — diminishing returns

Комбинации

Сочетайте с другими MCP — эффект x10

skillcompass-skill + skill-optimizer-skill

Two complementary tools: SkillCompass ranks, skill-optimizer drills into SKILL.md patterns

Use SkillCompass to pick the worst skill; use skill-optimizer to deeply analyze its SKILL.md.✓ Скопировано

skillcompass-skill + filesystem

Operate across the full ~/.claude/skills/ directory

Evaluate every skill in ~/.claude/skills/ and give me a CSV.✓ Скопировано

Инструменты

Что предоставляет этот MCP

Инструмент	Входные данные	Когда вызывать	Стоимость
skill-scoring	skill path(s)	Periodic audits	eval compute
weakest-link-id	bundle scores	After scoring	0
fix-suggestion	weak skill + critique	Before editing	0
before-after-eval	pre/post SKILL.md	After applying fixes	eval compute

Стоимость и лимиты

Во что обходится

Квота API: none beyond your LLM provider (evals use LLM calls)
Токенов на вызов: evals can be heavy — budget 20–100k tokens for a full bundle scan
Деньги: free, MIT
Совет: Run on one skill at a time during iteration; bundle runs only for audits

Безопасность

Права, секреты, радиус поражения

Хранение учётных данных: none at skill level

Исходящий трафик: none beyond your LLM provider

Устранение неполадок

Частые ошибки и исправления

Node errors on install

Ensure Node 20+; npm i inside the skill directory.

Проверить: node -v

Evals are inconsistent run-to-run

Fix the task seed and use a non-stochastic sample; record provider+model.

Альтернативы

SkillCompass в сравнении

Альтернатива	Когда использовать	Компромисс
skill-optimizer-skill	You want a single skill analyzed deeply rather than a bundle ranked	Depth over breadth
manual review	You have 1–2 skills total	Doesn't scale

Ещё

Ресурсы

📖 Читать официальный README на GitHub

🐙 Открытые задачи

🔍 Все 400+ MCP-серверов и Skills