菜单
Generates AGENTS.md file using the WHAT/WHY/HOW framework. Explores the codebase and produces a minimal (<100 line) context file with progressive disclosure.
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
Isolated Codex goal-loop skill that combines spec-to-plan, goal-loop-native phase planning, and gated phase-by-phase execution. Use when the user asks to turn a spec into a goal-loop plan, run a goal-loop plan, or package the spec-to-goal-plan, goal-loop-planner, and goal-loop flow as one Codex skill. Planner work uses GPT-5.4 high, implementation and gate-fix subagents use model 5.5 high, and review gates use model 5.4.
Autonomous phase-by-phase plan execution loop. Plan-subagent decomposes each phase into a dependency-tagged wave graph, parallel Opus sub-agents implement via the inlined orchestrator pattern (main thread plans/synthesizes, sub-agents do all edits), advisor + /code-review --effort extra-high + the inlined review-swarm protocol gate every phase, IMPLEMENTATION_NOTES.md captures off-spec decisions. Self-contained — orchestrator and review-swarm protocols ship inline as references, no third-party skill dependencies. Use when the user invokes /goal-loop <plan-path>, when /goal hook fires with a goal-loop directive, or when the user wants autonomous phase-by-phase execution of a markdown plan file. Pairs with goal-distill (which produces compatible plan prompts).
Distill the current session — conversation, decisions, corrections, repo state, and recent history — into a single paste-ready /goal prompt that drives an autonomous phase-by-phase loop (decompose → Plan-subagent on Opus → main-thread execute → re-measure → repeat) until the success criterion is met or a stop condition fires. Universal — works in any repo, writes no files. Copies the prompt to the clipboard on macOS. Use this whenever the user says "distill goal", "distill this session", "goal-distill", "reconstruct goal", "rebuild the spec", "rebuild the masterplan", "give me a /goal", "phase-by-phase to the goal", "loop until done", "iterate toward the goal", or "turn this into a /goal prompt I can paste later" — and prefer it even when the user doesn't name the skill but clearly wants this session compressed into a runnable prompt for a future agent.
Deep strategic analysis of a project to identify the single highest-leverage, most innovative addition. Use this whenever the user asks what to build next, what the most impactful improvement would be, what the "next big thing" is for their project, what's the smartest thing to add, or any question about strategic direction and priorities. Also trigger when the user seems stuck choosing between competing features, is evaluating what matters most, or asks open-ended questions like "what would you do with this project?" or "what am I missing?"
Enter orchestration-only mode where Claude plans, delegates to sub-agents, and synthesizes results instead of implementing directly. Use whenever the user says they want to act as tech lead, wants Claude to orchestrate, wants sub-agents to do the work, says things like "you orchestrate, I brainstorm", "don't code yourself", "delegate this", "break this up and dispatch it", or explicitly invokes /orchestrator. Also use when the user frames the session as brainstorming-to-spec-to-delegation and wants Claude to stay out of the implementation seat.
Parallel read-only multi-agent root-cause investigation for bugs, regressions, crashes, flaky behavior, or unexplained failures. Use when the user asks to investigate a bug, find the root cause, trace a regression, understand why something broke, or wants a ranked diagnosis with the fastest proof path without making code edits.
| name | init |
| description | Generates AGENTS.md file using the WHAT/WHY/HOW framework. Explores the codebase and produces a minimal (<100 line) context file with progressive disclosure. |
| argument-hint | [--reset] |
This skill is CLI-agnostic. It works with Claude Code, Codex, Amp, or any AI coding assistant.
Generate a minimal, high-impact AGENTS.md context file for this repository using the WHAT/WHY/HOW framework. See reference/AGENTS.md for the template, rules, and progressive disclosure specs.
Validate the result of every operation. If any step fails, halt and report the failure before continuing.
$ARGUMENTS
--reset: Regenerate everything from scratch -- overwrite AGENTS.md and all context files created by this skill, then re-run the full exploration.If $ARGUMENTS contains --reset:
.maestro/context/ files were created by this skill (not by maestro:setup). The skill-created files use snake_case names like building_the_project.md, running_tests.md, code_conventions.md, service_architecture.md, database_schema.md, etc. The maestro:setup files use kebab-case: product.md, tech-stack.md, guidelines.md, product-guidelines.md, workflow.md, index.md.maestro:setup files)../AGENTS.md if it exists.Read-only exploration. Do NOT ask the user for permission to explore -- just do it.
Search for .maestro/context/product.md.
If it exists, maestro:setup has been run. Read these files for pre-fill data:
.maestro/context/product.md -- purpose, users, features.maestro/context/tech-stack.md -- languages, frameworks, tools.maestro/context/guidelines.md -- coding conventions.maestro/context/workflow.md -- build/test methodologyStore findings as pre-fill. Do NOT ask questions the context already answers.
Regardless of whether maestro context exists, explore the codebase to discover or verify:
README.md, package.json / pyproject.toml / Cargo.toml / go.mod / build.gradle / pom.xml / Gemfile / composer.json (whichever exists).Makefile, justfile, Taskfile.yml, CI config (.github/workflows/*.yml, .gitlab-ci.yml), docker-compose.yml.CLAUDE.md if it exists -- extract any rules worth preserving../AGENTS.md if it exists -- note what it covers before overwriting..eslintrc*, prettier*, biome.json, ruff.toml, .rubocop.yml, clippy.toml, .editorconfig. Note what they enforce (used by template rules to avoid duplication).packages/, apps/, crates/, services/), database configs, API patterns.The agent decides what to read based on what it finds. This is exploration, not a rigid checklist -- adapt to the project.
Organize discoveries into these categories (internal notes, not output):
Use the template and rules from reference/AGENTS.md. The output file MUST be under 100 lines.
Use the progressive disclosure guidance from reference/AGENTS.md to create well-named files in .maestro/context/.
Create .maestro/context/ if it does not exist:
mkdir -p .maestro/context
Write ./AGENTS.md (overwrite if exists).
Write each progressive disclosure file to .maestro/context/.
Display summary:
AGENTS.md generated.
- ./AGENTS.md ({line_count} lines)
- .maestro/context/building_the_project.md
- .maestro/context/running_tests.md
{additional files as created}
Next steps:
- Review ./AGENTS.md and edit manually for accuracy
- /maestro:AGENTS.md --reset -- regenerate from scratch
Recommended workflow:
/maestro:setup -- Scaffold project context (run first)/maestro:AGENTS.md -- You are here. Generate AGENTS.md context file/maestro:new-track -- Create a feature/bug track with spec and plan/maestro:implement -- Execute the implementation/maestro:review -- Verify implementation correctness/maestro:status -- Check progress across all tracksAGENTS.md is an optional context enhancement that complements /maestro:setup. While setup creates project-level context files, this skill generates a codebase-oriented AGENTS.md with progressive disclosure. Run it after setup to give all agents richer context about the repository structure.