菜单
When the user approves a plan and wants to execute it. When the user says "execute the plan" or "let's go" or "start implementation". When a PLAN.md exists and is ready to be executed task-by-task.
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
When the user wants to add a new feature or capability. When the user says "I want to build" or "I'm thinking about". When requirements are unclear and need to be refined.
When the user says they are ready for review or wants to create a PR. When the user asks to check if code is ready for review. When the task or feature is complete and needs quality validation.
When the user is implementing a new feature or fixing a bug. When the user wants to write tests first before writing code. When the user is refactoring and needs to verify behaviour is preserved.
When the user wants to create an implementation plan. When the user says "let's implement this" or "create a plan". When a design has been approved and is ready to be broken into tasks.
Systematic code review workflow — evaluate changes for correctness, style, security, and test coverage before merging. When the user asks to review code or a pull request. When the user says "check code quality" or "look over this diff". When the user requests feedback on changes before merging.
Restart containers, prune images, tail logs.
| name | subagent-driven-development |
| description | When the user approves a plan and wants to execute it. When the user says "execute the plan" or "let's go" or "start implementation". When a PLAN.md exists and is ready to be executed task-by-task. |
| tools | ["read","exec","edit"] |
Without task isolation, a single agent accumulates context across all tasks — leading to context bleed, hallucinations from earlier tasks, and difficult-to-debug failures. This skill prevents that by dispatching a fresh subagent per task with two-stage review before committing each result.
prg plan generates a task listIsolate each task so earlier context cannot corrupt later work. Give the subagent: task description, relevant files, testing requirements — no knowledge of other tasks.
Let the subagent implement, write tests, and return code + test results before any review.
Reject before checking quality; a non-compliant implementation wastes quality review time. ✓ Does it match task requirements? ✓ Are all files modified as specified? ✓ Do tests pass? ❌ If no → reject, provide feedback, retry
✓ Follows project conventions? ✓ Handles edge cases? ✓ No obvious bugs? ✓ Code is readable? ❌ If no → request improvements
Commit after each task so failures are isolated and rollback is cheap.
git add <files>
git commit -m "Task N: brief description"
✅ Task 1: Setup database schema (3min) 🔄 Task 2: In progress...
Each task produces: committed code, passing tests, and a progress entry in the session log.
❌ Giving subagent entire codebase context ❌ Skipping Stage 1 review ❌ Continuing after failed tests ❌ Not committing after each task