菜单
Analyze and fix issues in the last completed task. Use when user reports errors after task was marked as done, when completion report had high confidence but result has issues, when AI missed obvious problems, or when post-task issues appear.
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
Test API endpoints with proper authorization including Basic Auth and session cookies. Use when testing curl requests, checking API responses, getting 401 Unauthorized, session expired errors, or verifying endpoint behavior after changes.
Implement accumulated improvements from improvements-backlog into AI rules and skills. Use when backlog has 5+ items, weekly rule update cycle, 2+ high priority improvements exist, before important releases, when same error repeated 3+ times, when user says implement improvements/process backlog/update rules from errors.
Fix bugs, errors, exceptions, and unexpected behavior. Use when debugging, fixing errors, handling crashes, or when user mentions bug, error, broken, not working, crash, exception, 500 error, Application error, white screen, TypeError, undefined is not a function, or unexpected behavior.
Analyze Claude Code prompt cache efficiency and costs. Use when user asks about cache performance, token usage, session costs, how much a session cost, wants to optimize prompt caching, when costs seem high, when context window is filling up, or mentions cache-analyzer/cache hits/token costs/session metrics/cost analysis.
Review code for quality, security, and maintainability. Use when reviewing pull requests, examining code changes, doing QA checks, architecture review, or CTO review. Also trigger for COMPLEX tasks affecting more than 5 files, API contract changes, or data migrations.
Create new skills, modify and improve existing skills, create or update AI rules. Use when creating rules, building skills, adding coding standards, updating .claude/rules/ or .claude/skills/, converting Cursor rules to Claude Code, measuring skill performance, running evals, or optimizing skill descriptions.
| name | fix-last-task |
| description | Analyze and fix issues in the last completed task. Use when user reports errors after task was marked as done, when completion report had high confidence but result has issues, when AI missed obvious problems, or when post-task issues appear. |
| allowed-tools | Read, Write, Bash, Grep, Glob |
| model | sonnet |
| user-invocable | true |
Analyze failure before fixing. Patching without understanding why you missed the problem guarantees you'll miss it again.
Apply 5 Whys — dig at least 5 levels deep to find why issues were missed. Surface-level answers like "I forgot" are not root causes — keep asking until you reach a systemic reason.
Create TODO list — one item per missed issue, critical ones first.
For each TODO:
[before]% -> [after]% — [reason]Open and read EVERY changed file. Run the Challenge (4 questions) again.
Record improvement to .cursor/data/improvements-backlog.md because patterns repeat without reflection — the same class of mistake will resurface unless codified.
Full DONE block: issues found, fixes applied, root cause, improvement recorded, final confidence.
For detailed analysis and review templates, see templates.md.