/ 目錄 / 演練場 / woodshed
● 社群 gaearon ⚡ 即開即用

woodshed

作者 gaearon · gaearon/woodshed

Create, run, rate, and iterate your Claude Skills — Dan Abramov's skill-workshop skill, a meta-loop for getting skills actually good.

A meta-skill for authoring other skills. Scaffolds a new skill, runs it on test cases, asks you to rate the output, and uses the feedback to iterate on the skill's prompts. Based on Dan Abramov's 'woodshedding' methodology — deliberate practice applied to skill authoring.

為什麼要用

核心特性

即時演示

實際使用效果

woodshed-skill.replay ▶ 就緒
0/0

安裝

作者儲存庫 · Claude Skills 官方文件

使用場景

實戰用法: woodshed

Author a new skill and iterate it to 'actually works'

👤 Skill authors who've noticed their first drafts don't survive real use ⏱ ~120 min advanced

何時使用: You have an idea for a skill but every time you write one, it degrades after 3 uses.

步驟
  1. Scaffold
    Use woodshed. Create a new skill called 'changelog-writer' that turns git logs into user-facing changelogs.✓ 已複製
    → skills/changelog-writer/ created with SKILL.md
  2. Create test cases
    Generate 5 diverse test inputs.✓ 已複製
    → tests/ folder with 5 sample git log fixtures
  3. Run and rate
    Run the skill on each. Show me the output; I'll rate.✓ 已複製
    → Outputs presented with rating slots
  4. Iterate based on ratings
    Based on my ratings, suggest prompt changes to SKILL.md. Don't apply yet.✓ 已複製
    → Proposed diff with rationale per change

結果: A skill that works on your real use cases, not just the demo.

注意事項
  • Overfitting to test cases — Hold out some tests; re-run after every iteration to catch regressions
搭配使用: filesystem

Improve a skill that's been disappointing you

👤 Anyone with a skill installed they're unhappy with ⏱ ~60 min intermediate

何時使用: You have a skill that half-works and you can't tell why.

步驟
  1. Characterize the failures
    Use woodshed. Here's a skill I installed. Here are 3 sessions where it underperformed. Diagnose.✓ 已複製
    → Pattern in failures — instruction gap, example gap, framing issue
  2. Propose fixes
    Propose minimal SKILL.md edits to address those patterns.✓ 已複製
    → Specific diffs

結果: A skill you can trust.

組合

與其他 MCP 搭配,撬動十倍槓桿

woodshed-skill + skill-fetch-skill

Find skills with skill-fetch, improve them with woodshed

Find a changelog-writer skill; install; run woodshed on it to tune for my repo.✓ 已複製
woodshed-skill + filesystem

Persist iteration history across sessions

Keep woodshed logs in skills/<name>/.woodshed/ for traceability.✓ 已複製

工具

此 MCP 暴露的能力

工具輸入參數何時呼叫成本
scaffold_skill skill name, goal New skill 0
run_tests skill + test set Evaluating 0
capture_rating rating + notes After each test 0
iterate_prompt skill + ratings After a rating round 0

成本與限制

運行它的成本

API 配額
None
每次呼叫 Token 數
High during iteration (many test runs)
費用
Free (Claude API usage applies)
提示
Keep test sets small — 5 tests beats 50 that never get reviewed.

安全

權限、密鑰、影響範圍

憑證儲存: None
資料出站: None from skill

故障排查

常見錯誤與修復

Iteration keeps changing but doesn't improve

Your test set is too narrow; add diverse cases

Skill works in woodshed but not in real use

Production prompts differ from test fixtures; add real-session logs to the test set

替代方案

woodshed 對比其他方案

替代方案何時用它替代權衡
Hand-tuning SKILL.md without loopSimple skills with obvious promptsSlower convergence; easier to miss regressions
claude-skills-vault-skillYou want ready-made skills, not authoring toolsDifferent goal

更多

資源

📖 閱讀 GitHub 上的官方 README

🐙 查看未解決的 issue

🔍 瀏覽全部 400+ MCP 伺服器和 Skills