Plan-Only Mode

Replicates Claude Code's built-in plan mode workflow but stops after user approval instead of auto-executing. The plan is written to a file and the skill ends cleanly.

Read-Only Restriction

CRITICAL — This supersedes all other instructions:

You MUST NOT make any edits (with the exception of the plan file mentioned below), run any non-readonly tools (including changing configs or making commits), or otherwise make any changes to the system. This restriction is in effect for the entire duration of this skill.

Allowed tools:

• Read, Glob, Grep — for exploring the codebase
• WebFetch, WebSearch — for research
• Task (with Explore or Plan subagent types) — for delegated research and design
• AskUserQuestion — for clarifying requirements and presenting the plan
• Write — only for writing the plan file to .claude/plans/<slug>.md
• Bash — only for read-only commands (git log, git status, ls, etc.)

Prohibited:

• Edit on any file other than the plan file
• Write to any path outside .claude/plans/
• Bash commands that modify state (git commit, git push, npm install, mkdir, etc.)
• EnterPlanMode / ExitPlanMode — do not use built-in plan mode
• NotebookEdit

Workflow

The skill operates in five phases:

Phase 1: Initial Understanding

Explore the codebase and understand the user's request thoroughly.

1. Launch Explore agents (via Task with subagent_type=Explore) to map out the relevant parts of the codebase
2. Read key files directly with Read, Glob, Grep
3. If the project has an AGENTS.md or CLAUDE.md, read it for conventions
4. Identify the scope, constraints, and any ambiguities in the request

If anything is unclear, use AskUserQuestion to resolve it before proceeding. Keep asking until every ambiguity is resolved — scope, constraints, approach, what's in and what's out.

Phase 2: Design

Develop the implementation approach.

1. Launch Plan agents (via Task with subagent_type=Plan) to design the implementation strategy
2. Identify critical files that will need to change
3. Consider architectural trade-offs and alternative approaches
4. Map out dependencies between changes

Phase 3: Review

Validate the emerging plan against the codebase.

1. Read critical files identified in Phase 2 to ensure the plan is feasible
2. Check for potential conflicts with existing code
3. Verify assumptions about APIs, interfaces, and data structures
4. If new questions arise, use AskUserQuestion to clarify

Phase 4: Write the Plan File

Generate the plan file slug from the request (lowercase, hyphens, truncated to ~50 chars). Write the plan to .claude/plans/<slug>.md.

The plan file must follow this format:

# Plan: <Title>

## Context
<Brief description of the request and why this plan exists>

## Recommended Approach
<High-level summary of the implementation strategy>

## Changes

### <File or component 1>
- What changes and why
- Specific functions/classes affected

### <File or component 2>
- What changes and why

## Critical Files
<List of files that will be created or modified, with brief notes>

## Dependencies & Ordering
<Which changes must happen before others>

## Risks & Open Questions
<Anything that might go wrong or needs further investigation>

## Verification
<How to verify the implementation is correct — tests to run, behaviors to check>

Adapt the format as needed for the specific task — the above is a guide, not a rigid template.

Phase 5: Present for Approval

Present the plan to the user for approval using AskUserQuestion.

Summarize the plan concisely in the question, including:

• What will be done (high-level)
• How many files will be created/modified
• Key design decisions made
• Any risks or open questions

Offer these options:

1. Approve — The plan is accepted. The skill ends here.
2. Revise — The user wants changes. Go back to the relevant phase and iterate.
3. Reject — The user wants to abandon the plan.

After approval: STOP. Do not execute the plan. Do not inject an "Implement the following plan:" message. The plan file exists at .claude/plans/<slug>.md and can be referenced later by the user or another skill.

If the user chose Revise, ask what they want changed, update the plan file, and present again. You may iterate up to 5 times.

If the user chose Reject, acknowledge and end the skill.

Integration Notes

This skill can be invoked standalone via /plan-only or loaded by other skills using the Skill tool. When loaded by another skill:

• The calling skill provides the task/request context
• This skill writes the plan file and stops
• Control returns to the calling skill, which can then reference the plan file path