grok-routing

name	grok-routing
description	When Claude Code should delegate to Grok vs handle work itself
user-invocable	false

When to call Grok

Prefer Grok (grok:* agents or slash commands)

• Substantial debugging after Claude is stuck
• Second-opinion implementation of a non-trivial change
• Best-of-N alternative approaches (--best-of-n)
• Risky edits that should land in a worktree (--worktree)
• Ambiguous architecture → /grok:plan then implement, or /grok:design
• Multi-PR delivery from a design doc → /grok:execute-plan
• Named multi-agent recipes → /grok:workflow
• PR CI/review babysitting → /grok:babysit
• Image/video generation (/grok:image, /grok:video)
• Documents (/grok:document --type pptx|pdf|docx)
• Structured or adversarial code review before shipping
• Long-running investigation better as a background job

Prefer staying in Claude

• Quick questions, renames, one-line fixes
• Tiny edits with obvious answers
• Pure conversation without repo mutation

Intent → mode

User intent	Prefer
Stuck on a bug / implement a fix	/grok:rescue (+ --worktree --check)
Unclear approach before coding	/grok:plan
Architecture / design doc + PR plan	/grok:design
Ship a design doc’s PR DAG	/grok:execute-plan (or --latest after design)
Multi-dimension structured fan-out	/grok:workflow
Ship quality on a branch/PR	/grok:review (+ --post-pending for GH)
Challenge design assumptions	/grok:adversarial-review
Watch / fix open PRs	/grok:babysit
Deck / PDF / Word	/grok:document
Brand stills / clips	/grok:image / /grok:video

Depth pipeline (multi-PR / ambiguous product work)

Prefer this sequence over a single giant rescue:

1. /grok:plan — explore + plan.md when the approach is unclear.
2. /grok:design — consensus design doc + PR Plan → artifacts under .grok-designs/.
3. /grok:execute-plan --latest (or explicit path) — implement the PR DAG in worktrees.
4. /grok:review / /grok:babysit — quality and CI/review loop.

Use /grok:workflow when you have a named multi-agent recipe (fan-out review dimensions, etc.), not ad-hoc parallel rescues.

Memory and agent profiles

• Long multi-session work: prefer --memory (or project GROK_MEMORY=1) so Grok can reuse decisions.
• Codebase map / investigation without edits: --agent explore or --read-only --sandbox read-only.
• Planning only: /grok:plan or --plan on task.

Parallel agents

When workstreams are independent, run multiple Grok agents at once:

Need	Agent / command
Fix / investigate	grok:grok-rescue or /grok:rescue
Plan	grok:grok-plan or /grok:plan
Design	grok:grok-design or /grok:design
Execute plan	grok:grok-execute or /grok:execute-plan
Workflow	grok:grok-workflow or /grok:workflow
Review	grok:grok-review or /grok:review
Babysit	grok:grok-babysit or /grok:babysit
Document	grok:grok-document or /grok:document
Image / video	grok:grok-media or /grok:image / /grok:video

How:

1. Split into independent prompts.
2. Spawn multiple Agent tool calls in the same turn with run_in_background: true, or multiple companion Bash calls with --background.
3. Track each job id via /grok:status.
4. Collect results with /grok:result <job-id>.

Do not serialize independent Grok work just because another job is running.

Command map

Need	Command
Fix / investigate	/grok:rescue
Plan mode	/grok:plan
Design doc	/grok:design
Execute PR plan	/grok:execute-plan
Workflow	/grok:workflow
Review	/grok:review
Challenge design	/grok:adversarial-review
Babysit PRs	/grok:babysit
Document	/grok:document
Image	/grok:image
Video	/grok:video
Sessions	/grok:sessions
Progress	/grok:status
Output	/grok:result
Handoff transcript	/grok:transfer