writing-plans

name	writing-plans
description	Use when you have a spec or design document and need to break it into a detailed implementation plan with right-sized tasks

Writing Plans

Overview

Write comprehensive implementation plans assuming the engineer has zero context for our codebase and questionable taste. Document everything they need to know: which files to touch for each task, code, testing, docs they might need to check, how to test it. Give them the whole plan, with each task sized to a single self-contained unit of change. DRY. YAGNI. TDD. Frequent commits.

Assume they are a skilled developer, but know almost nothing about our toolset or problem domain. Assume they don't know good test design very well.

Announce at start: "I'm using the writing-plans skill to create the implementation plan."

Save plans to: .plans/YYYY-MM-DD-<feature-name>.md

Task Sizing

Each task is one self-contained, single-PR-scope unit of change — what would naturally land as a single commit and be reviewable on its own. Do not decompose tasks into atomic steps like "write the test" / "run the test" / "implement" — that level of choreography belongs to the implementer subagent, not the plan.

A task is sized correctly when:

• Its acceptance criteria fit in 1–3 specific, verifiable bullets
• It changes one conceptual unit (one component, one feature slice, one refactor phase)
• It produces one commit at the end

Vary task size by complexity:

• Simple work — renames, version bumps, doc updates, applying an existing pattern, one-line config changes. Bundle related simple work into a single task. Don't create separate tasks for "add the constant" and "use the constant" — that's one task.
• Standard work — a new function with tests, a new endpoint following existing patterns, refactoring one module's internals, adding a config option. One task. This is the default size.
• Complex or risky work — auth, security, or data-integrity changes; new abstractions; cross-cutting refactors; anything where you couldn't draft acceptance criteria without thinking hard. Keep tasks tight and split aggressively. Touching more than ~5 files or introducing a new abstraction is a signal to split.

When in doubt, size up. The cost of treating something as more complex than it is = one extra review pass. The cost of bundling something risky with something simple = shipping a bug inside a too-large diff.

No fixed cap on tasks per plan — let the plan be as long as the work is. But if a task's acceptance criteria don't fit in 1–3 bullets, it's too big — split it.

Plan Document Header

Every plan MUST start with this header:

# [Feature Name] Implementation Plan

For Claude: REQUIRED SUB-SKILL: Use Skill(executing-plans) to implement this plan task-by-task.

Goal: [One sentence describing what this builds]

Architecture: [2-3 sentences about approach]

Tech Stack: [Key technologies/libraries]

---

Task Structure

Each task must follow this structure:

### Task N: [Component Name]

Files:

• Create: exact/path/to/file.ext
• Modify: exact/path/to/existing.ext:123-145
• Test: tests/exact/path/to/test.ext

Acceptance Criteria:

• [Specific, verifiable bullet — e.g., "function([]) returns None and does not raise"]
• [...]

Notes: [Non-obvious context, dependencies on prior tasks, gotchas]

Commit: <type>(<scope>): <description>

The implementer subagent handles the red-green-refactor cycle internally — your job is to specify what counts as done, not how to get there.

Documentation Task

Before writing tasks, scan the project's documentation files (README.md, CLAUDE.md, docs/, etc.) and identify which sections would become stale after the planned changes. Then:

• If docs need updating: Add a final task that updates the specific files and sections affected. This task follows the same structure as any other task — list the files, the sections to change, the new content, and a commit. It gets spec-reviewed and code-reviewed like everything else.
• If no docs need updating (e.g., pure internal refactor): Add a comment at the end of the plan: <!-- No documentation updates needed --> so it's a conscious decision, not an oversight.

Remember

• Exact file paths always — but always relative to the repo root, never absolute paths
• Never hardcode the repository's absolute path (e.g., /Users/alice/project) anywhere in the plan — plans may be executed in worktrees at different paths
• Never include cd /absolute/path commands — use relative paths or assume the working directory is the repo root
• Complete code in plan (not "add validation")
• Exact commands with expected output
• Reference relevant skills with @ syntax
• DRY, YAGNI, TDD, frequent commits

Execution Handoff

After saving the plan:

• Commit the plan document to git

Then ask the user if they want to execute using AskUserQuestion:

AskUserQuestion(
  questions: [{
    question: "Plan is ready. How would you like to proceed?",
    header: "Execute",
    multiSelect: false,
    options: [
      { label: "Execute with subagents (Recommended)", description: "Full isolation - best for complex plans or autonomous work" },
      { label: "Execute quickly", description: "Faster - does implementation and reviews in main context" },
      { label: "Don't execute", description: "Stop here - execute manually later" }
    ]
  }]
)

Based on selection:

Execute with subagents:

• REQUIRED SUB-SKILL: Use Skill(executing-plans)
• Dispatches subagents for implementation and reviews
• Best for complex plans or autonomous work

Execute quickly:

• REQUIRED SUB-SKILL: Use Skill(executing-plans-quickly)
• Does implementation and reviews inline in main context
• Best for simple plans or interactive sessions

Don't execute:

• Plan is saved for later execution
• User can invoke execution skills in any session