About map
A routing skill for finding which project skill fits a task, from requirements discovery through delivery and review. It directs the agent to the relevant skill or manual slash command and leaves execution to that skill.
Key features
- 01Routes tasks across delivery stages
- 02Lists debugging and research paths
- 03Covers setup and orchestration utilities
- 04Resolves overlaps by delivery order
- 05Identifies tasks with no matching route
Use cases
- 01Choosing a skill for an incoming task
- 02Finding a path for vague requirements
- 03Locating debugging or research workflows
- 04Finding setup and orchestration commands
Install
Copy the command and run it in a terminal.
$
npx skills add https:/ / github.com/ yeachan-heo/ oh-my-claudecode --skill mapAfter installing, type/mapin Claude Code to confirm it loaded.
Details
From the project READMEMap
You know something needs doing; you do not know which skill owns it. Find the job below and follow its route. Most routes call the Skill tool; a route marked manual gives you the slash command to invoke yourself because model invocation is disabled. The map routes and explains nothing twice — each routed skill owns its own doctrine.
The delivery loop
| You are here | Skill |
|---|---|
| External work arrived (issues, PRs, reports) | harbor |
| A non-engineer handed you a chat log as requirements | intent |
| The destination is foggy — the first three decisions cannot be stated yet | ask-navigator |
| Vague idea; questions must come before anything else | deep-interview |
| Requirements are clear; you need a plan | plan (multi-perspective consensus: ralplan) |
| Approved plan; decompose and ship end to end with human checkpoints | launch |
| One task; persistence until verified complete | ralph |
| Hands-off idea-to-code | autopilot |
| Work needs coordinated parallel workers | team |
| One approved task carried through verified code | execute |
| A design question prose cannot settle | loft |
| Finished work needs judgment | review |
| Something claims to be done | verify |
| Work is landing as a PR | pr |
| The run taught the environment something | refit |
On-ramps (off the main loop)
| Job | Skill |
|---|---|
| A bug needs its root cause | debug; competing-hypothesis tracing: trace |
| Tests should drive the implementation | tdd |
| The codebase is getting harder to change | architecture-survey |
| AI-generated slop needs cleanup | ai-slop-cleaner |
| An open question needs grounded answers | research; external docs: external-context |
| Knowledge should compound across sessions | wiki; durable project memory: remember |
| A recurring autonomous improvement loop | self-improve; durable goal ledger: ultragoal |
| A session needs reorientation across its pivots and outcomes | session-flow — run /oh-my-claudecode:session-flow manually |
| Visual work needs verdicts or diagrams | visual-verdict; diagram |
Utilities
| Job | Skill |
|---|---|
| Install, diagnose, configure | omc-setup; omc-doctor; configure-notifications |
| Manage skills themselves | skill (local); skillify (extract from session) |
| Inspect orchestration state | trace (also the tracing lane); graph (DAG runtime) |
| Cancel a mode | cancel |
Rules
- The map routes; it never executes. Each ordinary row resolves to one Skill-tool call — make it, then follow that skill's doctrine. For a manual-only route, show the user its slash command; do not try to call the Skill tool.
- When two rows both fit, the delivery loop above is the tiebreak: earlier legs come first.
- Phase boundaries (continue / re-enter from disk / hand off / compact) are defined in
launch's Context hygiene — the map does not restate them. - A job matching no row is itself information: say so plainly rather than picking the nearest-sounding skill.
Comments
Markdown supported