菜单
Break down a goal, process, or workflow into MECE components with Agent SDK mapping. Use when user says "decompose", "break down this process", "MECE analysis", "create a decomposition tree", or pastes a JSON/YAML/CSV workflow export to analyze.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Opt the current project into down-tier model delegation by installing a standalone .claude/rules/model-delegation.md rule, optionally with pre-shaped .claude/agents/ definitions (fast-executor, task-coder) and an optional agent-state feedback layer. The rule tells Claude to route well-specified data and coding tasks to a cheaper model in a subagent, keeping judgment-heavy work in the main loop. Use when the user says "set up model routing", "set up model delegation", "use cheaper models for subagents", "delegate to a lower-power model", or wants to remove or update that rule.
Pre-share scan for leaked secrets and privacy-sensitive content. Wraps simonw/scan-for-secrets for literal matching (with JSON/URL/HTML/backslash/unicode escape variants) and composes a ripgrep regex pass for shape-based leaks (other users' home paths, emails, IPv4, MAC, JWTs, API keys). Use when user says "scan for secrets", "check for leaked credentials", "pre-share scan", "redact home paths", "PII scan", "strip my username from transcripts", "scan logs before publishing", "check before sharing", "audit before commit", "find leaked API keys", "scan agent transcript".
Convert JavaScript and TypeScript projects to use bun instead of npm/yarn/pnpm. Use when user asks to "migrate from npm to bun", "convert to bun", "switch package managers", "replace npm", or "use bun instead". Includes conversion tables and lock file migration. Core conventions auto-loaded via SessionStart hook; invoke /dev-conventions:bun-tooling for full reference.
Audit project dependencies for known vulnerabilities using uv audit (Python) and bun audit (JS/TS). Use when user says "audit dependencies", "check for vulnerabilities", "security audit", "check deps", "are my packages safe", "CVE check", "dependency security", "vulnerability scan". Invoke with /dev-conventions:dep-audit.
Documentation conventions: last-updated dates, lowercase filenames, internal/ subfolder for non-shared docs, session logs, and documenting the "why". Use when creating or updating documentation, READMEs, or design docs. Invoke with /dev-conventions:doc-conventions. Triggers on "write docs", "update README", "document this", "add a design doc", "session log".
Detailed Python/uv conversion reference. Core conventions auto-loaded via SessionStart hook; invoke /dev-conventions:python-tooling for full conversion tables. Use when you need the complete uv command mapping, version pinning strategy, or lock file workflow.
| name | decompose |
| description | Break down a goal, process, or workflow into MECE components with Agent SDK mapping. Use when user says "decompose", "break down this process", "MECE analysis", "create a decomposition tree", or pastes a JSON/YAML/CSV workflow export to analyze. |
| metadata | {"author":"Fred Bliss","version":"0.4.1","last_verified":"2026-04-02"} |
Decompose a process, goal, or workflow into MECE (Mutually Exclusive, Collectively Exhaustive) components with dual output -- a human-readable tree and structured JSON mapping to Claude Agent SDK primitives.
/decompose <description of what to decompose>
Examples:
/decompose our customer onboarding process from sign-up to first value delivery/decompose the CI/CD pipeline for our microservices architecture/decompose quarterly financial close process/decompose then paste a JSON/YAML/CSV export from a workflow toolIf the user provides structured data (JSON, XML, YAML, CSV, or any workflow/process export), do NOT assume you understand the schema. Ask what the data represents, what key fields mean, and how entities relate. Present your interpretation back, let them correct it, repeat until agreed. For free-text, skip this.
Establish boundary, trigger, and completion criteria. If ambiguous, ask clarifying questions. Do not guess at scope boundaries.
Score candidate dimensions (temporal, functional, stakeholder, state, input-output) using the 4-criteria rubric from the mece-decomposer skill. Document the winner and rationale.
Produce 3-7 L1 components. Apply full MECE validation (see references/validation_heuristics.md).
Decompose each L1. Choose dimension per branch (may differ from L1). Decrease validation rigor with depth per the depth-adaptive schedule.
At each leaf, apply the co-occurrence heuristic. Classify atoms by execution type: agent, human, tool, or external.
Identify data, sequencing, resource, and approval dependencies between branches.
Map atoms to Agent SDK primitives per references/agent_sdk_mapping.md. Assign model tiers.
Run final structural checks. Compute ME/CE scores.
Process Name (orchestration type)
+-- Phase 1 (parallel)
| +-- [agent] Step A (~5m, sonnet)
| +-- [human] Step B (~2h, webhook)
+-- Phase 2 (sequential)
+-- [tool] Step C (~10s, tool_name)
+-- [agent] Step D (~1m, haiku)
Each line: label, execution type in brackets, estimated duration, model tier or integration method.
Full JSON conforming to references/output_schema.md. To validate structurally:
uv run mece-decomposer/skills/mece-decomposer/scripts/validate_mece.py <output.json>
After producing the JSON, call the mece-decompose MCP tool with the full JSON string to render the interactive tree in Claude Desktop, Cowork, or Claude.ai.
After decomposition:
/validate/interview/export