Plan Mode

Problem

When an agent starts by writing code, editing files, or running commands, it can drift from the actual goal, miss side effects, and create work that is hard to review. plan-mode forces the agent to produce a plan first and block implementation until the user explicitly approves the approach.

How It Works

The skill is triggered by /plan. It then outputs a structured plan:
- Task Understanding: goal, expected deliverable, success criteria, constraints, and open questions.
- Context Review: modules, dependencies, interfaces, existing patterns, and affected logic.
- Missing Information: key assumptions, with at most three focused questions when needed.
- Implementation Plan: 3-8 steps, each with action, reason, impact, and expected output.
- Risk and Validation: regression risks, compatibility, performance, security, maintenance cost, tests, and edge cases.

After the plan is printed, the agent stops and waits for approval such as go, approved, or “execute”. Once approved, it enters execution mode: follow the approved plan, pause on material deviations, and keep progress updates concise.

Fit and Limits

This is useful for cross-module changes, interface updates, refactors, and production work where impact review matters. For trivial one-line edits or when the user explicitly wants immediate execution, the planning step can feel unnecessary.