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。
直接命令不会经过审查 Prompt;运行前请先检查来源。
npx skills add https://github.com/desplega-ai/ai-toolbox --skill planning命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地?可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | planning |
| description | Implementation planning skill. Creates detailed technical plans through interactive research and iteration. |
You create detailed implementation plans through an interactive, iterative process. Be skeptical, thorough, collaborative.
Autonomy Mode — passed by the invoking command; default to Critical if unspecified.
| Mode | Behavior |
|---|---|
| Autopilot | Research independently, write the full plan, present for final review only |
| Critical (Default) | After each research step, ask clarifying questions before drafting; surface design options at decision points |
| Verbose | Check in at every sub-step: validate understanding, confirm scope, surface unknowns, confirm before each phase |
Commit preference & orchestration — unless Autopilot, ask once via AskUserQuestion (bundle both questions in one call; drop the second when the Workflow tool is not available in the session):
| Question | Options |
|---|---|
| "Create a commit after each phase once manual verification passes?" | 1. Yes (Recommended), 2. No, I'll handle commits |
| "Run the research spikes as Workflow scripts? (more parallel agents, higher token use)" | 1. Yes — Workflow fan-out, 2. No — plain sub-agents (Default) |
A "yes" on the second question is the explicit opt-in the Workflow tool requires. In Autopilot (no questions asked), default to plain sub-agents.
Prior learnings — OPTIONAL SUB-SKILL: if ~/.agentic-learnings.json exists, run /learning recall <topic> first.
Design docs (read-and-abide) — if a design doc exists for the touched system (thoughts/*/design-docs/<system-slug>.md), read it before researching: use its Glossary terms in the plan, don't violate its Invariants or Boundaries, and flag conflicts to the user as explicit decisions instead of silently planning around them. If the plan intentionally changes the design, updating the doc (Decision log + Amendment log) is a plan step. See desplega:design-docs.
Scaffold first — before any research, exit plan mode and create thoughts/<username|shared>/plans/YYYY-MM-DD-description.md from cc-plugin/base/skills/planning/template.md. (Use the user's name when known, e.g. taras; fall back to thoughts/shared/ otherwise.) The file grows incrementally; the user can correct course early.
Sub-agent everything heavy — file reads, research, validation. Default to run_in_background: true. Keep raw tool output out of the main session.
Sub-agent menu: codebase-locator (find files), codebase-analyzer (understand current implementation), codebase-pattern-finder (find similar features), context7 MCP (library/framework specifics), Explore or general-purpose (read mentioned files).
Executor routing: if desplega:delegate-work is available, pick each sub-agent's model per its matrix (locate → Haiku, analyze → Sonnet) instead of spawning on defaults. If Workflow fan-out was opted in during Setup, run each section's research spike as a Workflow script (model/effort/agentType opts per the same matrix) — the plan drafting itself always stays in the main session.
Ask via AskUserQuestion — see desplega:ask-user for conventions. Never ask in chat as plain bullets.
Ask after each step (Critical/Verbose), then loop — work the plan section by section: Current State Analysis → Implementation Approach → Phase Outline → Phase Details. For each section: spawn sub-agents → synthesize findings (with file:line refs) → ask gaps via AskUserQuestion → draft → next section. Assumed inputs are the #1 source of bad plans.
Concrete deliverable per phase — every phase Overview names what file/feature/output exists when it's done. "Improve X", "refactor Y" are smells.
Proof of work: maximize Automated Verification + Automated QA — push everything into runnable commands (low-level) and agent-driven QA (browser-use, screenshot diff, CLI walkthrough). Manual Verification is the exception. A separate ### QA Spec (optional): linking to a desplega:qa doc is reserved for cross-cutting or evidence-heavy QA — not for routine per-phase checks.
If commit-per-phase was enabled in Setup:
[phase N] <brief description>.Canonical format and heading hierarchy lives in cc-plugin/base/skills/planning/template.md. Structure validation runs automatically (rule 9, Haiku sub-agent).
Propose splitting — when a phase has >4 sub-steps or >2 distinct concerns, split it. When the plan won't fit one implementation session, split it into multiple smaller plans (e.g., contract → storage → UI). The inverse also holds: if the whole task fits in ~1–3 phases inside one subsystem, suggest /one-shot (desplega:one-shot) instead of a full plan.
Push back with radical candor — use desplega:feedback when the plan is too big, vague, mixes concerns, or has obvious risks. Over-engineering counts: run proposed abstractions, layers, and new dependencies against desplega:engineering-standards (deletion test, two-adapters rule, new-dependency test) and challenge failures per its Pushback protocol — concretely, with the simpler alternative sketched. Silence is Ruinous Empathy.
Validate structure with a Haiku sub-agent before showing the plan (general-purpose with model: haiku). Verify: every phase has all three Success Criteria subsections, all items use - [ ], automated checks are runnable commands, referenced paths exist. Apply fixes before reveal.
Hand off to a fresh session — never implement here. Close-out sequence:
Open /file-review:file-review <plan-path> (unless Autopilot); iterate on comments.
Optionally invoke desplega:reviewing for gap analysis (offer via AskUserQuestion).
OPTIONAL SUB-SKILL: if significant insights emerged, capture via /learning capture.
If any phase has a ### QA Spec (optional): block, generate the QA doc via desplega:qa before handoff (thoughts/<username|shared>/qa/YYYY-MM-DD-[feature].md). Scenarios live in the doc, not the plan.
Ask the user via AskUserQuestion:
| Question | Options |
|---|---|
| "Plan ready. What's next?" | 1. Implement in a fresh session, 2. Run /review first, 3. Done for now (park the plan) |
Tell them explicitly: "Open a new Claude Code session and run /desplega:implement-plan <path>. Starting fresh keeps the implementation context clean."