ams-planning
暂无简介
- License 许可
- MIT
- 最近更新
- 2026-11-07
- Stars
- ★ 0
- Forks
- 1
03 SKILL.md 正文
Planning
Overview
Agents execute well but plan poorly without structure. Planning is the discipline of decomposing a goal into sequenced, assignable steps before any agent acts.
In the AI Maestro model, planning produces an explicit plan: goal, steps, agent assignments, dependencies, checkpoints, and success criteria.
Good plans make multi-agent work reliable: each agent knows its task, its inputs, its outputs, and when to hand off.
Plan once, execute many: the plan is reviewable, resumable, and reusable.
When to use
- Complex tasks needing multiple agents or steps
- Agent work that keeps drifting or duplicating effort
- Coordinating handoffs between specialized agents
- Long-running work that must survive interruptions
- Reviewing agent strategy before burning execution budget
Core concepts
- Decomposition. Break the goal into steps small enough for one agent to complete confidently. Each step: clear input, clear output, verifiable done-ness.
- Sequencing and dependencies. Order steps by dependency: what must finish before what can start? Parallelize independent steps; sequence the dependent ones.
- Agent assignment. Match steps to agent capabilities: researcher for gathering, coder for building, reviewer for checking. Right agent, right step.
- Handoff contracts. Define exactly what passes between steps: file paths, data formats, key facts. Vague handoffs are where multi-agent work dies.
- Checkpoints. Human (or supervisor) review points before irreversible or expensive steps. Plans without checkpoints drift.
- Success criteria. Verifiable conditions for 'done' per step and overall. Without them, agents optimize for activity, not outcomes.
- Failure branches. For each risky step: what if it fails? Retry, fallback step, or escalate? Plans that assume success break on contact with reality.
- Plan as artifact. The plan lives in a file: readable, reviewable, resumable. Execution updates it; it never lives only in chat.
Practical workflow
- State the goal. One paragraph: outcome, constraints, non-goals. The plan serves the goal — get it crisp first.
- Decompose into steps. Break down until each step is agent-sized: clear inputs, outputs, and done criteria. Number them; note dependencies.
- Assign agents. Match each step to the best-suited agent type. Note required context each agent needs at handoff.
- Define handoffs. For each boundary: what artifact passes, in what format, where. Write the contracts explicitly.
- Add checkpoints and criteria. Review points before big steps; success criteria per step and overall. Make 'done' verifiable.
- Review the plan. Human reviews before execution: are steps right-sized? Dependencies correct? Risks covered? Fix now, cheaply.
- Execute with tracking. Run steps in order; update plan status as you go. Blocked steps get noted with cause, not silently skipped.
- Close out and learn. Verify success criteria; record what the plan got right/wrong. Feed lessons into the next plan.
Common pitfalls
- Skipping planning. Throwing agents at a complex goal with no decomposition. Activity without architecture wastes budget and time.
- Oversized steps. 'Research the market' as one step. Steps must be completable in one focused agent run with verifiable output.
- Vague handoffs. 'Pass the results to the next agent.' What results, in what format, where? Handoff vagueness compounds across steps.
- No checkpoints. Long agent runs with no review. Drift compounds silently; checkpoints catch it while it's cheap.
- Planning forever. Analysis paralysis: planning the plan. Timebox planning; a good-enough plan executed beats a perfect plan drafted.
- Ignoring dependencies. Parallelizing steps that actually depend on each other. Dependency errors surface as rework — map them upfront.
- No failure branches. Assuming every step succeeds. One failed step with no fallback stalls the whole plan.
- Plan in chat only. Plans in scrollback can't be tracked, resumed, or reviewed. The plan is a file, always.
04 同仓库更多 Skills
web-scrape
Turn web pages into LLM-ready markdown via the Firecrawl API (renders JS, anti-bot handling). Use when browser_open/fetch fails on JS-heavy pages, or when you need clean markdown/structured extraction from a site. Requires FIRECRAWL_API_KEY.
design-taste-frontend
tasteskill v2, the anti-slop frontend skill. Art direction for websites, distilled from a study of 247 curated gallery sites (One Page Love, Awwwards Sites of the Day). Deals researched directions (idea, composition, type, palette, photographs, section blueprints) from a library, enforces a strict no-slop house style with a real-browser inspector, and runs a look-and-fix loop on screenshots. Use this whenever you build, design, redesign, restyle, polish or review a landing page, marketing site, portfolio, product page, one-pager, hero or any public-facing web page, in any stack (HTML/CSS, React, Next.js, Vue, Svelte, Astro, Tailwind) — also when the user says a page looks generic, templated, "AI-made", sloppy, or asks to make it look designed, premium, award-level or less like every other site. Not for dashboards or app screens.
test-driven-development
Use when implementing any feature or bugfix, before writing implementation code
systematic-debugging
Use when encountering any bug, test failure, or unexpected behavior, before proposing fixes
requesting-code-review
Use when completing tasks, implementing major features, or before merging to verify work meets requirements
brainstorming
You MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation).