跳转至

Subagent Cookbook — 15 个复制粘贴就能用的派遣 recipe

繁體中文 | 简体中文 | English

📋 这是什么Stage 5.5 教你 subagent 是什么概念,这份 cookbook 教你今天就能用——15 个场景,每个都包含“该用哪个 subagent + 完整 prompt 模板(复制即可用)+ 何时不用”。

⚠️ 第一次看?先读 Stage 5.5 的“可派遣的 subagent 有哪些”+“decision table”两节——理解“什么是 subagent”“Claude Code 内置有哪些”之后再来查 recipe。


怎么读这份 cookbook

每个 recipe 都用同样的 4 段结构:

内容 为什么有这段
场景 你今天工作会遇到的具体场景 从“我有 X 问题”找到 recipe,而不是从“我想用 subagent”找
Subagent 用哪一个(Claude Code 内置名称) 直接 copy 名字,不用想
Prompt 模板 复制粘贴即可用的指令文字 不用自己想怎么写
何时不用 比用 subagent 更好的替代方案 避免“大材小用”

💡 怎么实际派遣 subagent:在你的 Claude Code 终端对话框里,直接输入(或粘贴)prompt 模板——就这样。Claude 看到指令,会自动通过 Task tool(内部派遣机制)找到对应 subagent 跑,跑完后向主 session 回报一段摘要。不需要 slash command,不需要特殊语法

📌 subagent ≠ slash command/agents 是列表命令,不是调用 subagent 的方式;派遣 subagent 直接打对话 prompt 文字即可。完整对比表(subagent vs skill / vs slash command / description router)见 Stage 5.5 §易混淆观念厘清


先确认你有哪些 subagent 可用

在你的 Claude Code session 里跑 /agents 这个指令,会列出全部当前可用的 subagent(内置 + plugin + 自定义)。

Claude Code 默认内置 7 个 subagent:general-purpose / code-reviewer / Explore / Plan / frontend-developer / claude-code-guide / statusline-setup(截至 2025-11,可能会变动)。

📌 每个内置 subagent 的功能说明 + “遇到 X 任务用 Y”decision table 的 canonical 在 Stage 5.5 §可派遣的 subagent 有哪些——本 cookbook 聚焦“怎么派遣”的 15 个 recipe,内置清单与选用逻辑以 Stage 5.5 为准。

💡 如果 /agents 列表跟这份 cookbook 不一致:表示你装了 plugin,或 Claude Code 版本不同。recipe 名字对不上时,找最接近的 subagent 用就好(例如没有 Explore,用 general-purpose 也能跑搜索)。

确认完当前可用清单之后,就可以照下面 15 个 recipe 派遣了——找到符合你场景的 recipe,把 Prompt 模板复制到 Claude Code 对话框即可


15 个 Recipe

Recipe 1: 写完一批新 code,要 commit 前的 review

场景:你刚写了 ≥ 50 行新 code,想 commit 但怕有 bug / security 问题没看到

Subagentcode-reviewer

Prompt 模板

Review the staged changes in this repo (git diff --cached).
Focus on: (1) security issues (hardcoded secrets, SQL injection, XSS),
(2) error handling gaps, (3) missing tests, (4) violations of CLAUDE.md
conventions. Per-category PASS/FAIL + concrete fix for each issue with
file:line reference + overall verdict (APPROVE / REQUEST CHANGES).

何时不用:< 20 行的 typo / formatting fix(直接 git diff 自己扫就好,不用花 subagent token)


Recipe 2: 进入新 repo,不知道该从哪个 file 开始读

场景:clone 完一个别人的 repo,README.md 讲不清楚程序入口在哪,不想随便挖

SubagentExplore

Prompt 模板

Map the entry points and core structure of this codebase. Report:
(1) main entry script(s) — what file gets run first,
(2) core module organization — top 3-5 directories and what each does,
(3) test directory layout,
(4) any "where to start reading" guidance in docs/.
Under 300 words with file paths.

何时不用:你已经知道要看的 file 路径(直接用 Read tool)


Recipe 3: 设计 refactor / migration plan(先想清楚再动手)

场景:要重构一个大 module 或做 framework migration,不确定步骤,想先有 plan 再开始

SubagentPlan

Prompt 模板

Design a step-by-step plan to refactor <module-name> (currently a single
file with tight coupling) into clear, testable components with explicit
interfaces. Include: (1) phased breakdown (≤ 5 phases), (2) which files
touched per phase, (3) what tests gate each phase, (4) rollback strategy
if a phase fails. Don't write code — just the plan.

何时不用:refactor 只动 1-2 个 file(直接动手,不需要 plan overhead)


Recipe 4: 多文件 / 跨 locale parity 审查

场景:你改了 zh-TW + zh-Hans + en 三个 mirror,想确认三边内容一致

Subagentcode-reviewer

Prompt 模板

Review the staged diff for cross-locale parity across the 3 locale
variants of <file-stem> (.md / .zh-Hans.md / .en.md). Check: (1) same
section structure (## headers match), (2) same table row counts, (3)
same required terms present in each locale, (4) locale conventions
correct (zh-Hans uses "" not 「」; en uses English). Report per-file
PASS/FAIL.

何时不用:只改 1 个 locale(没有 parity 问题)


Recipe 5: 多 source fact-check / research

场景:要写一段引用,不确定多个 source 讲的是不是同一件事,需要交叉验证

Subagentgeneral-purpose

Prompt 模板

Fact-check this claim: "<claim>". Search for: (1) primary source / official
docs / paper, (2) 2-3 independent secondary sources, (3) any contradictions
or version differences. Report: confirmed / contradicted / nuanced, with
direct quotes and URLs. Under 400 words.

何时不用:你已经有 1 个权威 source(直接读那一个就好)


Recipe 6: 找某个 symbol / function 在哪定义

场景:codebase 里到处都在 import parse_config,想知道实作在哪个 file

SubagentExplore

Prompt 模板

Find where `<symbol-name>` is defined in this codebase. Report: (1) the
file:line where it's defined, (2) what it does (1-line summary), (3) which
files import / use it (top 5). Use Grep, not full file reads.

何时不用:你已经用 IDE 的 "Go to definition" 可以跳过去(IDE 比 subagent 快)


Recipe 7: 比较多篇 paper / source 的 claim 是否一致

场景:写文献回顾,3 篇 paper 对同一件事说法不同,不知道该信谁

Subagentgeneral-purpose

Prompt 模板

Find and compare 3 independent sources on the topic <topic>. For each
source, capture: title + author + year + URL. Then report: (1) what
they agree on, (2) where they disagree (with direct quotes), (3) which
is most recent / authoritative, (4) suggested phrasing if you had to
summarize the consensus. Flag if a key source is behind a paywall and
unreadable.

何时不用:3 篇都是同一个作者 / 同 lab(没有“多 perspective”可比)


Recipe 8: Release commit security audit

场景:要发 v1.0 / 升 major version,想最后做一次安全扫描

Subagentcode-reviewer

Prompt 模板

Pre-release security audit for <release-tag> (use git log to find the
relevant commit range since last release). Check: (1) hardcoded
credentials / API keys, (2) eval() / exec() / shell injection risks,
(3) deprecated dependencies with known CVEs, (4) any auth / session
handling changes since last release, (5) public API surface changes
(breaking? documented?). Per-category PASS/FAIL + remediation per
finding.

何时不用:patch release(hotfix)只动 1 行(人工扫 < 1 分钟)


Recipe 9: 评估架构变动的 blast radius

场景:想改一个 base class / 共用 utility,不确定会影响多少 file

SubagentPlan

Prompt 模板

Assess the blast radius of changing <component>. Report: (1) direct
dependents (which files import this), (2) indirect dependents (transitive
imports), (3) tests that gate the change, (4) suggested rollout order
(safest → riskiest), (5) feature flag / kill switch strategy if available.
Don't make the change yet — just the impact analysis.

何时不用:只影响 1 个 file 内部(没有 blast radius 概念)


Recipe 10: Spawn 并行 subagent 跑同一任务 × 多目标

场景:4 个 branches file 都要做同样的“academic-style audit”,想一次跑 4 个 parallel

Subagentgeneral-purpose × N(同时 spawn 多个)

💡 怎么“同时 spawn 多个”:在 Claude Code 对话框内,连续输入下面的 prompt 4 次(每次换 <file-path>)。Claude Code 会自动并行跑,不需要等第一个 subagent 结束才能输入第二个——这就是“parallel spawn”的实际操作方式。

Prompt 模板(每次换 <file-path>):

Audit `<file-path>` for academic-style issues: (1) over-engineering
jargon without first-use explanation, (2) clarity (long sentences,
vague pronouns), (3) unsupported % claims, (4) persona-fit (wrong
technical level for the file's stated target audience — see the
file's banner / intro callout for who it's for). Report 4-category
PASS/FAIL + fix-list with line numbers. Under 500 words.

何时不用:1 个目标就够(直接调用 1 次,不要 over-orchestrate)


Recipe 11: 找跨 repo 的相似 implementation

场景:你在某个 repo 看到一个 pattern,想知道另一个 repo 有没有类似实作

SubagentExplore

Prompt 模板

In <repo-path>, search for code patterns similar to: <description of
pattern, e.g., "retry decorator with exponential backoff">. Report:
(1) up to 5 candidates with file:line, (2) brief description of each,
(3) which is most idiomatic for this codebase's style.

何时不用:你有具体 function 名要找(用 Recipe 6 / Explore 加精确 grep 更快)


Recipe 12: LLM-as-judge eval(structured PASS / FAIL)

场景:跑了 100 个 test case,要评每个 output 是否符合 spec,不想人工看

Subagentgeneral-purpose

Prompt 模板

Evaluate the agent outputs in <eval-file.jsonl> against the spec stated
in the file's first 5 lines (or in <spec-file.md>). For each row: PASS /
FAIL with 1-sentence reason. Aggregate: pass rate + top-3 failure modes.
Output as structured JSON: {"case_id": "...", "verdict": "PASS|FAIL",
"reason": "..."}.

何时不用:< 5 个 case(自己看更快);evaluation 涉及主观判断(LLM judge 不可靠)


Recipe 13: UI component design / accessibility audit

场景:写了一个 React component,想确认 ARIA + keyboard nav + responsive 都做对

Subagentfrontend-developer

Prompt 模板

Audit <component-file> for: (1) ARIA roles + labels (screen reader
compatibility), (2) keyboard navigation (tab order, Enter / Esc / arrows
behavior), (3) responsive breakpoints (mobile 360px / tablet 768px /
desktop 1280px), (4) color contrast (WCAG AA), (5) touch target size
(≥ 44px). Report per-category findings + fixes.

何时不用:纯后端 / CLI 工具(没有 UI 可 audit)


Recipe 14: 问 Claude Code feature 怎么用

场景:忘了 hooks 怎么设、忘了 slash command 的 frontmatter 字段,想查文件

Subagentclaude-code-guide

Prompt 模板

How do I <specific Claude Code feature question, e.g., "configure a
PreToolUse hook to block dangerous bash commands">? Show: (1) minimum
config in settings.json, (2) example hook script (Python or shell),
(3) 1 common gotcha, (4) where in the official docs to read more.

何时不用:你已经写过几次(直接看自己 ~/.claude/settings.json 范例更快)


Recipe 15: 写 React form validation 逻辑

场景:要做一个 sign-up form,要 email format / password strength / real-time validation

Subagentfrontend-developer

Prompt 模板

Implement a React sign-up form with: (1) email format validation
(real-time, debounce 300ms), (2) password strength meter (≥ 8 chars,
mixed case, digit, symbol), (3) inline error messages with ARIA
live region, (4) submit button disabled until valid. Use <library, e.g.,
React Hook Form + Zod>. Include: component code, validation schema,
1 happy-path test, 1 error-path test.

何时不用:非 React stack(用 Vue / Svelte 对应的 subagent,或 general-purpose


Recipe 索引(按 subagent type 找)

不确定该用哪个 subagent?从任务类型反查:

Subagent Recipes
code-reviewer 1(pre-commit review)/ 4(cross-locale parity)/ 8(release security audit)
Explore 2(new codebase)/ 6(find symbol)/ 11(cross-repo similar code)
Plan 3(refactor plan)/ 9(blast radius)
general-purpose 5(fact-check)/ 7(multi-paper compare)/ 10(parallel multi-target)/ 12(LLM-as-judge eval)
frontend-developer 13(a11y audit)/ 15(React form)
claude-code-guide 14(Claude Code feature 查询)

何时 NOT 该用 subagent

Subagent 不是免费的——每次派遣会烧 token,也有延迟。下面 4 种场景不该用 subagent,自己跑更划算:

  1. 任务 < 5 分钟可自己完成 — 大材小用;subagent overhead 不划算
  2. 结果需要逐步 user feedback — subagent 是“派出去、跑完回报一次”,不能中间问你;要逐步确认的任务直接在主 session 跑
  3. 任务需要主 session 的 context memory — subagent 是独立 context window,看不到主 session 前面的对话;要用“我们刚刚讨论的 X...”这种 reference 的任务,不适合
  4. 任务涉及 conversation-level judgment — 像“这个 architecture 决策该怎么权衡”,需要对话协作,不适合丢给 subagent

💡 判断的快速办法:如果任务可以用“写一份完整 brief 给陌生人接手做”描述清楚 → 适合 subagent;如果需要“我跟你讨论一下”→ 留在主 session。


接下来