Plan Skill

Create a structured, checkable implementation plan before writing any code.

When to use

• Task requires 3 or more steps
• Task touches multiple files or layers
• Task involves an architectural decision
• Any time the path forward is not immediately obvious

Protocol

1. Read mandatory context first

• claude/tasks/todo.md — current task state (are we resuming?)
• claude/tasks/lessons.md — relevant past lessons
• claude/architecture.md — project structure
• claude/rules.md — golden rules

2. Understand the full scope

• What is the exact goal? What does done look like?
• What layers does this touch (schema, service, API, tests, docs)?
• What already exists that is relevant?
• What are the failure modes?

3. Write the plan to claude/tasks/todo.md

Format:

# Task: <Clear, concise title>

**Goal:** <One sentence. What is done when this is complete?>
**Branch:** <git branch name, if applicable>
**Loaded docs:** <which claude/*.md files you read>

## Steps
- [ ] 1. <First action — specific file or command>
- [ ] 2. <Second action>
- [ ] 3. <Third action>
...

## Review Checklist
- [ ] Tests pass
- [ ] Linter passes
- [ ] No new warnings
- [ ] PR description ready (if applicable)

## Notes
<Any constraints, risks, or decisions to track>

4. Validate the plan (optional — for complex tasks)

Spawn a plan-challenger subagent:

"Spawn plan-challenger to review this plan. Here is the plan: [paste plan]"

5. Get approval for non-trivial plans

Present the plan to the user. Wait for "proceed" or revisions before implementing.

Rules

• NEVER start implementing before the plan is written
• NEVER mark a step complete without proof (test result, output, etc.)
• If something goes wrong mid-execution: STOP, re-read the plan, update it, then continue