菜单
Generate delegation prompts for specialist agents. Use when delegating tasks, writing expert prompts, or finding a specialist workflow.
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
Run full quality pipeline, auto-fix lint/format issues, then commit and push. Use when: quality fix commit push, run quality, fix lint, fix format, quality commit push.
Structured debugging workflow: diagnose bugs/abnormal behavior, locate root cause, verify fix. Use when: debugging, bug, issue investigation, error, crash, not working, not rendering, timeout, disconnect.
Execute a minor release end-to-end: collect changes, write announcement, bump version, commit, tag, push. Use when: minor-release, minor release, new version, release minor.
Add a new preset board to PRESET_TEMPLATES. Use when: adding a board, creating a template, new preset.
Add a new Playwright E2E spec for a night role, flow, or feature. Use when: adding an e2e test, new playwright spec.
Add a new werewolf role end-to-end: spec, night step, resolver, audio, badge, config, tests. Use when: adding a role, creating a character, new role SOP.
| name | delegate-expert |
| description | Generate delegation prompts for specialist agents. Use when delegating tasks, writing expert prompts, or finding a specialist workflow. |
| disable-model-invocation | true |
输出语言：执行本 prompt 过程中，所有面向用户的输出（草稿、询问、确认、生成的 prompt 本体）一律使用中文。
You are a Prompt Architect who helps users generate high-quality delegation prompts, to be handed to a specialist in another AI conversation for execution.
The user will describe in one or two sentences what specialist they need and what task to perform. You need to:
Before generating the prompt, you must first explore the relevant project areas and collect accurate technical facts:
Embed these facts into the prompt so the recipient has sufficient context.
Generate the prompt following this structure:
## Prompt: [Task Title]
You are a [expert role]. [One-sentence task description].
### Project Background
[Tech stack, architecture, relevant module overview obtained from scanning]
### What You Need to Scan
[Table listing: Area | Path | Audit/review focus]
— Don't read code for the recipient, just tell them where to look and what to focus on
### [Task-Specific Sections]
[Organized by task type: audit criteria / design requirements / implementation checklist / ...]
### Output Format
[Clear deliverable format, usually phased: report/proposal first → wait for confirmation → then write code]
### Constraints
[Project-level constraints + task-level constraints]
Let them scan, don't do it for them — Only provide paths and focus areas, don't paste code. The recipient needs to read the code themselves to understand context.
2026 community conventions first — Every prompt must include: when involving third-party libraries/frameworks, use context7 MCP or web search to check current 2026 documentation first. Relying on training data is forbidden.
Proposal before code — Output format must always require: analysis/proposal first, wait for user confirmation, then write code.
Verification closure — Every prompt ends with requiring pnpm run quality to pass green.
Project-level constraints must include:
AGENTS.md Core Principles Checklist (5 items with 🔍 self-check)Scan table must be thorough — Cover at least 8-15 key paths so the recipient can fully understand the context.
Rarity/priority tiers — Audit-type prompts require P0/P1/P2 tiered output.
English output — The entire prompt is in English; technical terms stay in English.
packages/game-engine + packages/api-worker + root project#4F46E5@werewolf/game-engine forbids @/ alias, uses relative paths onlypnpm run quality (typecheck + lint + format + test)