SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsclaude-review

claude-review

Agents & models

View repository

Updated 10-15 19:58

About this skill

Claude Cross-Review Workflow

Principles

  • Those who write don't review: The main Claude focuses on writing code and does not self-review boundary conditions
  • Those who review don't write: Review subagents focus on finding problems and do not modify code
  • Parallel multi-dimensional: Three subagents review simultaneously from different angles, improving coverage
  • Loop until clean: If there are issues, fix them; after fixing, review again, until all subagents report no new issues

Trigger Timing

After completing feature development, bug fixes, or refactoring, run this workflow before git commit .

Process

Step 1: Confirm the scope of work

git diff --stat
git diff HEAD

List the changed files and confirm the review scope. Save the full output of git diff HEAD for later use (all three agents need it).

Step 2: Dispatch three review subagents in parallel

In the same message, use the Agent tool to simultaneously start the following three subagents (subagent_type=general) and let them run in parallel. Each agent is responsible for only one review dimension.

Agent 1 — Logic and boundary conditions

你是一位专注逻辑与边界条件的代码审查员。只输出问题列表，不评价整体质量。没有问题则输出：NO_ISSUES

【审查维度】
- 边界条件：空数组/空字符串、零值、极大/极小值、索引越界
- 逻辑漏洞：条件判断遗漏、逻辑短路错误、状态机缺失状态
- 控制流：提前返回遗漏、循环条件错误、递归终止条件缺失

【代码变更】
<git diff HEAD 完整输出>

【输出格式】
问题 #1：[逻辑/边界] [文件路径:行号]
  描述：...
  风险：...
  修复建议：（思路即可，不写代码）

Agent 2 — Type safety and error handling

你是一位专注类型安全与错误处理的代码审查员。只输出问题列表，不评价整体质量。没有问题则输出：NO_ISSUES

【审查维度】
- 类型安全：nil/null 解引用、类型断言失败、接口不匹配、类型转换溢出
- 错误处理：忽略 error 返回值、错误信息丢失、panic 未恢复
- 空值安全：可选值未检查、指针解引用前未判空

【代码变更】
<git diff HEAD 完整输出>

【输出格式】
问题 #1：[类型/错误] [文件路径:行号]
  描述：...
  风险：...
  修复建议：（思路即可，不写代码）

Agent 3 — Concurrency and resource safety

你是一位专注并发与资源安全的代码审查员。只输出问题列表，不评价整体质量。没有问题则输出：NO_ISSUES

【审查维度】
- 并发安全：数据竞争、死锁、未加锁的共享状态、goroutine/协程泄漏
- 资源管理：文件/连接未关闭、内存泄漏、defer 使用错误、事件监听器未清理
- 幂等性：重复调用是否安全、重试逻辑是否正确

【代码变更】
<git diff HEAD 完整输出>

【输出格式】
问题 #1：[并发/资源] [文件路径:行号]
  描述：...
  风险：...
  修复建议：（思路即可，不写代码）

Step 3: Consolidate issues

Wait for all three agents to finish, merge the outputs, deduplicate, and sort by priority:

Priority: logic flaws > type safety > concurrency safety > boundary conditions > resource management

If all three agents output NO_ISSUES → skip to Step 6

Step 4: The main Claude fixes item by item

Fix one by one in priority order:

  • Update progress after each fix:已修复 #N or 跳过 #N（说明原因）
  • Do not over-fix: only change the issues pointed out by the subagents, and do not casually refactor unrelated code

Step 5: Dispatch review again

Repeat Steps 2-3 and check:

  • Whether the fix correctly resolved the original issue
  • Whether the fix introduced new issues

If there are new issues → return to Step 4

If all are NO_ISSUES → continue to Step 6

Step 6: Verification

# 根据项目类型选择合适的验证命令
# lint / typecheck
# build / compile
# test

After all pass, proceed to Step 7. If verification fails → fix and return to Step 2.

Step 7: Commit

git add <相关文件>
git commit -m "$(cat <<'EOF'
<commit message>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
EOF
)"

Output Format

Report at the end of the workflow:

## Claude Cross-Review 完成报告

**审查轮次**：N 轮
**发现问题**：N 个（逻辑 N + 类型 N + 并发 N）
**已修复**：N 个
**跳过**：N 个（含原因）
**验证结果**：全部通过 / 部分失败

已提交：<commit hash> <commit message>