Capability Router
commands ~236 tokens updated 2026-08-04
/capabilities - a decision map for CCGM's overlapping command/skill clusters. Answers "which one do I use?" for research, review, planning/execution, debugging, and knowledge/memory. A tight always-on rule points at the on-demand command so the full map costs zero idle tokens.
Tags
README
Capability Router
CCGM ships many commands and skills, and several clusters overlap (three research entry points, half a dozen review tools, four planning/execution commands). This module answers the recurring question: which one do I use?
What it installs
- A tight always-on rule (
rules/capability-map.md) with the most-confused one-liners and a pointer to the full map. Kept deliberately small so it costs almost nothing when idle. /capabilities- an on-demand command that prints the full decision map. The bulky map lives here, not in an always-loaded rule, so the token cost is paid only when you ask.
/capabilities [cluster]
Prints a decision map for the overlapping clusters. Pass a cluster name to print just that section:
/capabilities
/capabilities research
/capabilities review
/capabilities plan
/capabilities debug
/capabilities knowledge
Clusters covered
| Cluster | Picks |
|---|---|
| Research | /research (no deps) vs /deepresearch (Exa MCP) |
| Review | scope-drift, /ce-review, pr-review-toolkit, document-review, editorial-critique, design-review, adrev, /resolve-pr-feedback, built-in /review |
| Planning & execution | /xplan, /xplana, /etp, /mawf |
| Debugging | /debug vs the systematic-debugging methodology rule |
| Knowledge & memory | /reflect (personal) vs /compound (team) vs session-history |
The map notes which entries depend on external tooling and skips anything not installed in your setup.
Manual Installation
mkdir -p ~/.claude/commands ~/.claude/rules
cp commands/capabilities.md ~/.claude/commands/capabilities.md
cp rules/capability-map.md ~/.claude/rules/capability-map.md
Dependencies
None. The router documents other modules but does not require them - it tells you to skip entries you have not installed.
Will install
| Path | Action | Target | Type |
|---|---|---|---|
commands/capabilities.md | → | commands/capabilities.md | command |
rules/capability-map.md | → | rules/capability-map.md | rule |
Dependencies
No dependencies.
Required by
No other module depends on this one.
Included in presets
Install this module
Agent prompt
Recommended for agent users -- hands the whole install off to your assistant.
Fetch https://cd23a9be.ccgm-site.pages.dev/modules/capability-router.md and install this module into my Claude Code setup.
Native plugin marketplace
One command via the native plugin marketplace -- additive, does not merge settings.json.
claude plugin install capability-router@ccgm
The marketplace path is additive, not a replacement: it installs commands, agents, and skills as native plugin components, but it does not perform the bash installer's deep settings.json merge, and it does not write the always-loaded global CLAUDE.md context. Rules are only injected via an opt-in SessionStart hook rather than being auto-loaded. Use the bash installer when those pieces matter to you.
Files
rule (1)
rules/capability-map.md
# Capability Router Several installed commands and skills overlap. When unsure which to use, run `/capabilities` for the full decision map. The most-confused picks: - **Research:** `/research` (no deps, web/GitHub/Reddit) → `/deepresearch` (needs Exa MCP key). - **Code review:** `/ce-review` (full orchestrated PR review) → `scope-drift` skill (intent-vs-diff check, run first). - **Plan/spec review:** `document-review` (before execution). **Prose review:** `editorial-critique`. **Visual review:** `design-review`. - **Plan then execute:** `/xplan` (interactive) or `/xplana` (autonomous) to PLAN → `/etp` to EXECUTE a ready plan/issue. - **Debug a failure:** `/debug` (run the workflow). The `systematic-debugging` rule is the always-on methodology it follows. - **Knowledge:** `/reflect` (personal, this machine) vs `/compound` (team, committed to `docs/solutions/`). Do not invent a command. If no installed capability fits, say so.
command (1)
commands/capabilities.md
--- description: Print the CCGM capability map - which command/skill to use among overlapping clusters (research, review, planning/execution, debugging, knowledge). allowed-tools: Read, Glob, Grep argument-hint: "[cluster: research | review | plan | debug | knowledge]" --- # /capabilities - Which Command Do I Use? CCGM ships overlapping capabilities. This is the decision map. If `$ARGUMENTS` names a cluster, print only that section; otherwise print the whole map. Some entries below may not be installed in this setup - check `~/.claude/commands/` and `~/.claude/skills/` and skip anything missing. ## Research | Use | When | Notes | |-----|------|-------| | `/research` | Default. Broad web research with no setup. | Parallel agents over WebSearch, WebFetch, GitHub CLI, Reddit. **Zero external dependencies.** | | `/deepresearch` | You want semantic/neural search depth and have an Exa key. | Fans queries out via the **Exa MCP server** (requires `claude mcp add` + Exa API key). Synthesizes from full page contents, not snippets. | Decision: no API key or want it to just work → `/research`. Need higher-quality semantic retrieval and have Exa set up → `/deepresearch`. ## Review | Use | When | Scope | |-----|------|-------| | `scope-drift` (skill) | At the **start** of any PR review or before claiming a task done. | Compares stated intent (PR body, plan, TODOs) against the actual diff. Run this first, then a quality review. | | `/ce-review` | Full code-quality review of a PR/diff. | Orchestrates tiered reviewer personas (correctness, testing, maintainability + conditional security/performance/reliability/api-contract/migrations) plus an adversarial lens. Confidence-gated autofix. Modes: interactive/autofix/report-only/headless. | | `pr-review-toolkit` | You use the external pr-review-toolkit plugin. | Adds scope-drift + Fix-First output (AUTO-FIXED vs NEEDS INPUT) on top of that plugin. | | `document-review` | Reviewing a **plan/spec/requirements doc** before it ships to execution. | 7 lenses: coherence, feasibility, product, scope-guardian, design, security, adversarial. Not for code. | | `editorial-critique` | Reviewing **long-form prose** (essays, blog posts, reports). | 8 passes: prose craft, AI-tell detection, argument, conciseness, data, structure, impact, grammar. | | `design-review` | Reviewing the **visual design of a web page**. | Screenshots at 3 viewports; passes on spacing, typography, responsive, hierarchy, a11y, consistency. | | `adrev` (skill) | You want to **attack** a plan/doc/PR/idea, not grade it. | Separate agent steelmans the case against, hunts failure modes. Plan targets get findings folded in automatically. | | `/resolve-pr-feedback` | Existing PR has **review comments to resolve**. | Fetches unresolved threads, clusters them, applies unambiguous fixes, replies + resolves; batches taste questions. | | built-in `/review` | Quick built-in code review when none of the above is installed. | Claude Code's native command; no CCGM orchestration. | Decision: code → `/ce-review` (after `scope-drift`). Plan/spec → `document-review`. Prose → `editorial-critique`. UI → `design-review`. Want adversarial pressure → `adrev`. Resolving existing PR comments → `/resolve-pr-feedback`. ## Planning & Execution | Use | When | |-----|------| | `/xplan` | Interactive, human-in-the-loop planning of a new project/feature. Interview → research → tech-stack/scope sign-off → plan → reviews → execute. | | `/xplana` | Same pipeline as `/xplan --autonomous`: full depth, **no mid-flow prompts**, presents the finished plan at one final gate. | | `/etp` | You already have a **ready plan or GitHub issue(s)** and want to drive it to done with parallel agents, adversarial PR review, and follow-ups. `etp` executes; it does not research or write the plan. | | `/mawf` | You have **unstructured feedback** to turn into discrete GitHub issues, then spin up parallel agents to implement them. | Decision: still figuring out *what* to build → `/xplan` (guided) or `/xplana` (hands-off). Plan/issue ready to build → `/etp`. Pile of raw feedback to triage into issues → `/mawf`. ## Debugging | Use | When | |-----|------| | `/debug` | You are actually fixing a bug/error/failure. Runs a structured reproduce → hypothesize → instrument → diagnose → fix → verify workflow (delegates to a more capable model). | | `systematic-debugging` (rule) | Always-on methodology, not a command. It is the discipline `/debug` follows; you do not invoke it directly. | Decision: fixing something → `/debug`. The rule just enforces root-cause-first behavior in every session. ## Knowledge & Memory | Use | When | |-----|------| | `/reflect` | Capture a **personal** learning after a task. Writes the JSONL learnings store on this machine (confidence-scored, decaying). Never leaves your machine. | | `/compound` | Capture a **team-shared** learning. Writes a structured doc to `docs/solutions/` in the repo, committed and reviewed; `/xplan` and reviews later re-inject it as grounding. | | `session-history` (skill) | Search **prior session transcripts** for what was tried/failed/decided. Usually invoked by `/compound`, `/xplan`, `/debug` - not run standalone. | | memory MCP (if configured) | Cross-session structured memory via an MCP server, separate from CCGM's learnings store. | Decision: just for me → `/reflect`. For the team/repo → `/compound`. "What did a past session do here?" → `session-history`.