The single most important skill in this arsenal. Enforces disciplined agentic work.
When to Use
- User asks for a plan or explicitly says "plan mode"
- Task is larger than a single obvious edit
- Risk of breaking changes, architecture impact, or many files
- User says "don't write code yet"
Core Behavior
-
Immediately enter or confirm Plan Mode mindset.
- Explicitly state: "Entering Plan Mode. I will not edit files until you approve a plan."
-
Investigate first (read-only where possible).
- Use subagents for broad investigation if the scope is large.
- Run
grok inspectif configuration or skills may be relevant. - Gather evidence: relevant files, existing patterns, tests, risks.
-
Produce a high-quality plan with these sections:
- Goal (one sentence)
- Current state (what exists today)
- Proposed changes (numbered, smallest possible steps)
- Files likely to change
- Risks and mitigations
- Verification method (tests, manual steps,
grok inspect, etc.) - Estimated diff size / complexity
-
Present the plan and STOP.
- Do not proceed to implementation.
- Ask for approval: "Does this plan look good? Any changes before I proceed?"
-
After approval:
- Execute one step at a time where possible.
- Re-verify after each meaningful chunk.
- Use other skills (tdd-intelligence, git-discipline, etc.) as appropriate.
Subagent Usage
For large investigations, spawn 2–4 subagents with clear charters (e.g., "backend impact", "frontend impact", "test surface", "security surface") then synthesize.
Success Criteria
- User has explicitly approved the plan (or a revised version).
- All subsequent edits are traceable to the approved plan.
- Verification steps from the plan were actually performed.
Anti-Patterns to Avoid
- Jumping into edits "just to explore".
- Vague plans ("I'll refactor the auth layer").
- Plans that are really just implementation notes.
Example Trigger Phrases
- "Plan the migration to the new MCP protocol"
- "Use Plan Mode to add the arena runner feature"
- "Don't edit anything yet, just plan the security hardening"