Monkey Skills
Skill Development Conventions
File Paths
- SKILL.md 內引用 bundled files 時使用相對路徑(相對於 skill 目錄)
- Good:
checklists/security-checklist.md,protocols/code-brainstorming.md - Bad:
domain-teams/skills/code-team/checklists/security-checklist.md - 原因:Claude Code 提供 Base Path,bundled files 從 skill 目錄相對解析
Two-Layer Spec
- PRODUCT-SPEC.md(planning-team 擁有)— 跨域:商業 + 設計 + 技術方向
- TECH-SPEC.md(code-team 擁有)— 技術:模組設計 + 資料流 + 介面定義
- TECH-SPEC.md 應 reference PRODUCT-SPEC.md 的內容
Skill Structure
- 每個 team skill 是自包含目錄(SKILL.md + protocols/ + checklists/ + rubrics/ + standards/)
- SKILL.md body 控制在 ~6,000 tokens 以內(約 4,500 words);Anthropic 官方建議 <500 lines,本 repo 改用 token 計量因為行數密度差異大
- Domain knowledge 用目錄慣例 + 描述性檔名路由,不用靜態清單
- Reference files 從 SKILL.md 直接引用(one level deep,不巢狀)
Quality Gates
- 四級系統:SELF / MUST / SHOULD / MAY
- Gate 定義明確指定檔案路徑(相對路徑)
- Verdict 約束內嵌於 PASS_WITH_NOTES 定義,不另開段落
Agent Behavioral Rules
- worker:produces artifacts, does NOT produce gate verdicts
- evaluator:produces verdicts, does NOT modify artifacts
- Knowledge access is open(行為限制,非閱讀限制)
Agent Launch Convention
- 傳遞 檔案路徑 給 agent(不是檔案內容)
- Agent 用 Read 工具自行讀取資源
- 路徑在 SKILL.md 中用相對路徑定義,launch 時解析為絕對路徑
- Worker Input Contract: Resource Paths → Task → Input
- Evaluator Input Contract: Resource Paths → Artifact → Requirements
Cross-Plugin Delegation Contract
首例:investing-toolkit:investment-memo-writer → domain-teams:investing-team
規則:
- Delegation = pass paths + structured seed context — 不傳遞 file content,不內嵌分析結果
- Delegation target receives full authority — 被委派的 skill 自行載入 standards、執行 gates、產生 verdict;委派方不干涉
- Data layer stays in toolkit, analysis layer stays in domain-teams — investing-toolkit 只負責 data fetch + pipeline orchestration;investing-team 負責分析、primary-source anchoring、gate enforcement
- Gate verdicts flow back — delegation target 的 gate 結果(PASS / NEEDS_REVISION)回傳給 orchestrating skill,不被 swallowed
- Cross-plugin path resolution — 委派時使用 plugin name + skill path(e.g.
domain-teams:investing-team),不使用檔案系統絕對路徑
Pattern(investing-toolkit → domain-teams):
investing-toolkit skill
→ data-fetcher agent (I/O only)
→ domain-teams:{team} skill (analysis + gates)
→ domain-teams:docs-team (formatting, optional)
禁止:
- 不可在 toolkit skill 內自行執行 investing-team 的 gate logic(避免 gate bypass)
- 不可把 domain-teams standards 複製到 toolkit skill(避免 drift)
- data-fetcher agent 不可做分析(I/O only)