메뉴
L0/L1/L2 three-layer context loading protocol — reduces token consumption during /resume boot
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Use when about to claim work is complete, fixed, or passing, before committing or creating PRs - requires running verification commands and confirming output before making any success claims; evidence before assertions always. Upgraded with AC-first workflow, pre-flight fast-fail, and structured Judge verdicts (inspired by opslane/verify).
Guide for creating Claude Code skills following Anthropic's official best practices. Use when user wants to create a new skill, build a skill, write SKILL.md, or needs skill creation guidelines. Provides structure, naming conventions, description writing, and quality checklist.
Pattern for progressively refining context retrieval to solve the subagent context problem
Fast iteration with two-stage review (spec compliance, then code quality)
Multi-agent coordination SOP derived from Anthropic's internal coordinator architecture. Use when orchestrating parallel workers, dispatching subagents, planning multi-step implementations, or when tasks require research→synthesis→implementation→verification workflow. Triggers: 'coordinate', 'dispatch workers', 'parallel agents', 'multi-agent', 'fan out', 'orchestrate'.
Use to run multiple subagents concurrently on independent tasks
| name | layered-context |
| description | L0/L1/L2 three-layer context loading protocol — reduces token consumption during /resume boot |
Inspired by OpenViking (ByteDance) L0/L1/L2 architecture. Adapted for MUSE's pure Markdown zero-dependency design.
Full-loading all .muse/*.md files during /resume wastes tokens when the Agent only needs one role's context. The layered approach loads minimum context first, then deepens on demand.
| Layer | Token Budget | Content | When to Load |
|---|---|---|---|
| L0 | ~100 tokens | One-line HTML comment at top of each .muse/*.md | Always — scan ALL role files |
| L1 | ~2K tokens | Full role file content | On demand — only the CURRENT role's file |
| L2 | Unbounded | memory/*.md + code files + docs | On demand — grep search when needed |
Every .muse/*.md file MUST have an L0 comment as the first line:
<!-- L0: v2.10.1 | P0=竞品技术吸收, P1/P2全清, QA PASS, S036已接收 -->
<!-- L0: and -->)version | priorities | status<!-- L0: v2.10.1 | P0=竞品技术吸收(mem0/OpenViking), P1全清, QA 10/10 PASS -->
<!-- L0: 9/9渠道已发, Show HN暂缓, S040梗图排期中, Stars=2 -->
<!-- L0: 最近QA全PASS(10/10 v2.3), 无待修FAIL, QA清洁状态 -->
/resume [role]
│
├─① Read CLAUDE.md + MEMORIES.md (constitutional layer, always)
│
├─② Scan ALL .muse/*.md L0 lines (grep "<!-- L0:" .muse/*.md)
│ → Get one-liner status of every role in ~400 tokens total
│
├─③ Deep-read CURRENT role's .muse/*.md (L1, full file)
│ → Only the file matching /resume [role]
│
├─④ Scan memory/ for unfinished items (L2, on demand)
│ → grep 🔲 and [ ] in recent memory files
│
└─⑤ grep strategy.md for 🟡 directives (L2, on demand)
→ Only if non-strategy role
Agent receives a question/task
│
├─ Can answer from L0? → Answer immediately
│ (e.g., "What version is MUSE?" → L0 has it)
│
├─ Need role details? → Load L1 (full role file)
│ (e.g., "What's the P0 task?" → need build.md details)
│
└─ Need historical context? → Load L2 (memory/grep)
(e.g., "What did we decide about X last week?" → grep memory/)
The /bye workflow updates L0 as part of Step 3.5 (role file sync):
If the L0 comment's version doesn't match the latest tag, the /resume boot should flag it:
⚠️ L0 stale: build.md says v2.10.1 but latest tag is v2.11.0