planner（planner 效率）

Planner

Create structured, orchestrator-ready plans for multi-task projects.

Source: Adapted from am-will's codex-skills workflow patterns Pairs with: task-orchestrator skill for follow-on implementation

Quick Start

Load the full planner prompt from prompts/planner.md and follow its process:

1. Phase 0: Clarify requirements (ask up to 5 targeted questions)
2. Phase 1: Research & understand the codebase
3. Phase 2: Create detailed plan with sprints, tasks, acceptance criteria
4. Phase 3: Subagent review of the plan
5. Phase 4: Return the plan in markdown and save a file only if the user asked for a persisted artifact

Key Principles

Task Atomicity

Each task must be:

• Atomic and committable — small, independent pieces of work
• Specific and actionable — not vague
• Testable — include tests or validation method
• Located — include file paths and code locations

Bad vs Good Task Breakdown

❌ Bad: "Implement third-party sign-in"

✓ Good:

• "Add sign-in config to environment variables"
• "Install and configure the required authentication package"
• "Create sign-in callback route handler in src/routes/auth.ts"
• "Add the sign-in button to the login UI"

Sprint Structure

Each sprint must:

• Result in a demoable, runnable, testable increment
• Build on prior sprint work
• Include clear demo/verification checklist

Plan Template

# Plan: [Task Name]

**Generated**: [Date]
**Estimated Complexity**: [Low/Medium/High]

## Overview
[Brief summary of what needs to be done and the general approach]

## Prerequisites
- [Dependencies or requirements that must be met first]
- [Tools, libraries, or access needed]
- [Tooling limitations, e.g., browser relay/CDP restrictions]

## Sprint 1: [Sprint Name]
**Goal**: [What this sprint accomplishes]
**Demo/Validation**:
- [How to run/demo this sprint's output]
- [What to verify]

### Task 1.1: [Task Name]
- **Location**: [File paths or components involved]
- **Description**: [What needs to be done]
- **Perceived Complexity**: [1-10]
- **Dependencies**: [Any previous tasks this depends on]
- **Acceptance Criteria**:
  - [Specific, testable criteria]
- **Validation**:
  - [Test(s) or alternate validation steps]

### Task 1.2: [Task Name]
[...]

## Sprint 2: [Sprint Name]
[...]

## Testing Strategy
- [How to test the implementation]
- [What to verify at each sprint]

## Potential Risks
- [Things that could go wrong]
- [Mitigation strategies]

## Rollback Plan
- [How to undo changes if needed]

Hand-off

Once plan is ready, hand off to the parallel-task worker:

Please run parallel-task.md against my-plan.md

Or invoke directly:

"Run all unblocked tasks in plan.md using parallel subagents. Keep looping until all tasks are complete."

Safety Boundaries

• Do not save a plan file unless the user asked for one or the surrounding workflow explicitly needs a persisted artifact.
• Do not assign overlapping write scopes to parallel tasks without calling out the conflict.
• Do not invent dependencies, validation steps, or completion status that the repo context does not support.
• Do not turn a planning request into implementation work unless the user explicitly asks to move from planning to implementation.

Files

• prompts/planner.md — Full planner agent prompt
• prompts/parallel-task.md — Parallel task worker prompt

Both are based on am-will's codex-skills prompts.