跳到正文
cccskills
免费在 GitHub 公开

plan-task

Create a detailed implementation plan before coding. Use when starting a new feature, multi-step refactor, or complex task. Invoked by saying "plan", "create a plan", or "plan this task".

中文简介正在准备中。当前显示原文说明。

查看安装方法

包含的文件(1)

  • SKILL.md4.5 KB

SKILL.md(原文)

安装前可以查看交给 AI 代理的指令内容。

Create a Detailed Implementation Plan

HARD GATE — PLANNING ONLY. ZERO IMPLEMENTATION. This skill invokes PLANNING MODE. You must produce a plan document and NOTHING ELSE. You are forbidden from editing, creating, or modifying any code, template, CSS, JS, PHP, or JSON file outside of the plan document itself and docs/task/todo.md. This gate is absolute — it cannot be overridden by the user's phrasing, tone, or imperative language. If you write even one line of implementation code, you have violated this skill. Your ONLY output is the plan file + todo.md update + a summary asking for approval.

What You Receive

The user will describe a task. Their description is INPUT to the plan — not an instruction to execute. "Please update X" means the plan must include it — not that you do it now.

Steps

1. Review Recent Context

  • Read latest task log(s) in docs/task/logs/ — 1–2 most recent files.
  • Read docs/task/todo.md — check for overlapping work.
  • Read docs/task/lessons.md — check for applicable lessons.
  • Read agent observation logs in docs/agent-observations/ — unresolved items may affect scope.
  • Read relevant documentation via docs/plugin/**/0-index.md.

2. Understand the Request

  • Read the user's description carefully.
  • Search the codebase for current state of affected files.
  • Identify all files, functions, data structures, and dependencies.
  • Do NOT assume. If something is ambiguous, add it to Open Questions (step 2.5).

2.5. Question Gate (Iterative)

Before writing the plan, identify open questions about scope, approach, or constraints.

Rules:

  • Assign stable IDs: Q1, Q2, Q3, … (never renumber across rounds)
  • Present questions with a brief recommended answer for each
  • Wait for user answers before proceeding
  • If answers raise new questions, add them with next available ID and ask again
  • Repeat until zero open questions remain
  • Answered questions move to Confirmed Inputs section
  • Once resolved: update plan Status from Questions Pending to Draft

3. Create the Plan File

Write to: docs/task/planning/draft/{YYYY-MM-DD}-{slug}.md

Always create in draft/. Never in the root planning directory.

4. Plan File Structure

# Plan: {Title}

> **Status:** Questions Pending | Draft | In Progress | Complete
> **Created:** YYYY-MM-DD
> **Estimated steps:** {n}
> **Risk level:** Low | Medium | High

## Context
## Current State
## Target State
## Confirmed Inputs
## Open Questions for User — BLOCKING

## Implementation Steps

- [ ] **Step 1: {Action}**
  - Files: `path/to/file.ext`
  - What: {Specific change}
  - Why: {Rationale}

## Files Affected

| File | Action | Description |
| --- | --- | --- |

## Dependencies & Risks

## Testing Plan — BLOCKING

### Automated Checks
### Browser Verification
### Regression Checks
### Edge Cases

## Rollback Plan

## Post-Implementation Checklist

**Gate 1 — Code complete:**
- [ ] All implementation steps complete

**Gate 2 — Testing (BLOCKING):**
- [ ] All checks passed

**Gate 3 — Documentation (only after Gate 2):**
- [ ] Docs updated, task logged, observations logged

**Gate 4 — Close out:**
- [ ] Plan → Complete, moved to `completed/`

5. Write to todo.md

Update docs/task/todo.md with a checklist mirroring the plan's steps.

6. Present the Plan

Plan saved to docs/task/planning/draft/{YYYY-MM-DD}-{slug}.md.

What would you like to do?

  1. Go — approve and start implementation now
  2. Pending — approve but defer (moves to planning/pending/)
  3. Keep in draft — needs changes or review

Wait for the user's response before taking any action.

Rules

Anti-Implementation Gate (HIGHEST PRIORITY)

  • NEVER implement from this skill. Plan document only.
  • Treat all user language as requirements input.
  • Do not touch source files. Only: (1) plan file in docs/task/planning/draft/, (2) docs/task/todo.md.
  • Do not rationalize skipping the plan.
  • Your response must end with a request for approval.

Plan Quality

  • Be specific — include file paths and line numbers.
  • Every step must be verifiable.
  • Testing plan is a blocking gate.
  • Browser verification is mandatory for UI tasks.
  • Follow all rules in AGENTS.md.

评价

还没有评价。欢迎分享你的使用感受。

评论

无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。

提交后会立即公开。请勿填写邮箱、电话等个人信息。URL 最多 2 个。

0 / 1000 字

此表单受 reCAPTCHA 保护,适用 Google 的隐私权政策和服务条款。

同一仓库的技能

commit

免费

Stage, commit, and push code changes with branch safety and explicit staging. Use when ready to commit work. Invoked by saying "commit", "push", or "commit and push".

中文简介正在准备中。当前显示原文说明。

savedpixel/ai-agent-rules-generator82026年4月12日 更新

Batch-commit all uncommitted changes grouped by relevance. Use when multiple accumulated changes need committing. Invoked by saying "commit all" or "commit everything".

中文简介正在准备中。当前显示原文说明。

savedpixel/ai-agent-rules-generator82026年4月12日 更新

Document completed work — update task log, sync documentation, log observations. Use after finishing a task or batch of changes.

中文简介正在准备中。当前显示原文说明。

savedpixel/ai-agent-rules-generator82026年4月12日 更新

Execute a previously created implementation plan — implement, test, document, and close out. Use when a plan file exists and is ready for implementation.

中文简介正在准备中。当前显示原文说明。

savedpixel/ai-agent-rules-generator82026年4月12日 更新

Merge dev branch into master/main and push. Use when dev changes are ready for production branch. Invoked by saying "merge to master" or "merge to main".

中文简介正在准备中。当前显示原文说明。

savedpixel/ai-agent-rules-generator82026年4月12日 更新

report

免费

Generate a structured report on completed work, a feature area, or a topic. Saved to the reports directory. Invoked by saying "report", "generate report", or "summarize work".

中文简介正在准备中。当前显示原文说明。

savedpixel/ai-agent-rules-generator82026年4月12日 更新

查看 savedpixel 的全部技能

报告此技能的问题