AgentSkillsCN

multi-ai

根据给定请求启动多AI管道。引导完成计划->审查->实施->审查工作流。

SKILL.md
--- frontmatter
name: multi-ai
description: Start the multi-AI pipeline with a given request. Guides through plan -> review -> implement -> review workflow.
allowed-tools: Read, Write, Edit, Bash, Glob, Grep

Multi-AI Pipeline Orchestrator

You are starting the multi-AI pipeline. Follow this process exactly.

Reference Documents

First, read the standards that guide all reviews:

  • skill/multi-ai/reference/standards.md - Coding standards and review criteria

Step 1: Clean Up Previous Task

Remove old .task/ directory if it exists:

bash
rm -rf .task
mkdir -p .task

Step 2: Capture User Request

Write the user's request to .task/user-request.txt.

Step 3: Create Initial Plan

Write .task/plan.json:

json
{
  "id": "plan-YYYYMMDD-HHMMSS",
  "title": "Short descriptive title",
  "description": "What the user wants to achieve",
  "requirements": ["req1", "req2"],
  "created_at": "ISO8601",
  "created_by": "claude"
}

Step 4: Refine Plan

Research the codebase and create .task/plan-refined.json:

json
{
  "id": "plan-001",
  "title": "Feature title",
  "description": "What the user wants",
  "requirements": ["req1", "req2"],
  "technical_approach": "Detailed how-to",
  "files_to_modify": ["path/to/file.ts"],
  "files_to_create": ["path/to/new.ts"],
  "dependencies": [],
  "estimated_complexity": "low|medium|high",
  "potential_challenges": ["Challenge and mitigation"],
  "refined_by": "claude",
  "refined_at": "ISO8601"
}

Step 5: Sequential Plan Reviews

Run reviews in sequence. Fix issues after each before continuing:

  1. Invoke /review-sonnet

    • Read .task/review-sonnet.json result
    • If needs_changes: fix issues in plan, update .task/plan-refined.json
  2. Invoke /review-codex

    • Read .task/review-codex.json result
    • If needs_changes: fix issues and restart from step 5.1
    • If approved: continue to implementation

Step 6: Implement

Invoke /implement-sonnet

This skill will:

  • Read the approved plan from .task/plan-refined.json
  • Implement the code
  • Add tests
  • Output to .task/impl-result.json

Step 7: Sequential Code Reviews

Run reviews in sequence. Fix issues after each before continuing:

  1. Invoke /review-sonnet

    • Read .task/review-sonnet.json result
    • If needs_changes: fix code issues
  2. Invoke /review-codex

    • Read .task/review-codex.json result
    • If needs_changes: fix issues and restart from step 7.1
    • If approved: continue to completion

Step 8: Complete

Write .task/state.json:

json
{
  "state": "complete",
  "plan_id": "plan-001",
  "completed_at": "ISO8601"
}

Report success to the user with:

  • Summary of what was implemented
  • Files changed
  • Tests added

Important Rules

  • Follow this process exactly - no shortcuts
  • Fix ALL issues raised by reviewers before continuing
  • If codex rejects, restart the review cycle from sonnet
  • Keep the user informed of progress at each major step

State Files Reference

FilePurpose
.task/user-request.txtOriginal user request
.task/plan.jsonInitial plan
.task/plan-refined.jsonRefined plan with technical details
.task/impl-result.jsonImplementation result
.task/review-sonnet.jsonSonnet review output
.task/review-codex.jsonCodex review output
.task/state.jsonPipeline state

Reference Directory

PathPurpose
skill/multi-ai/reference/standards.mdReview criteria and coding standards
skill/multi-ai/reference/schemas/JSON schemas for structured output