Skill Router

Skill Router

Purpose

Scan all installed skills once, build a lightweight in-session index, then route subsequent requests to the most relevant skill — but only when a skill would materially improve the response. No overhead on conversational or trivial requests.

Two modes:

• Suggest mode (default): proposes matching skills ranked by relevance, waits for confirmation
• Auto mode: routes silently and appends *(via skill-name)* to routed responses

Activation

/skill-router on — suggest mode (default)

/skill-router on --auto — auto mode

Both run the same activation steps:

Step 1 — Scan installed skills

find ~/.claude/skills -name "SKILL.md" | sort

For each file found, extract the following fields (in order of preference):

1. name: from YAML frontmatter
2. description: from YAML frontmatter
3. triggers: or trigger: from YAML frontmatter (if present)
4. If any field is missing, infer it from the first 20 lines of the file

Skip skill-router itself. Skip files that cannot be read.

Step 2 — Build the index

Construct an in-memory table (session-only, never written to disk):

SKILL INDEX
-----------
<skill-name> | <trigger keywords, comma-separated> | <one-line description>
...

Also initialize an in-memory session usage counter:

SESSION_USAGE = {}   ← empty dict, incremented each time a skill is invoked

Example row:

geo-audit | geo, seo, audit, AI search | Full website GEO+SEO audit with parallel subagent delegation.

Rules:

• Extract trigger keywords from the trigger, triggers, or tags frontmatter fields
• If none exist, infer 3–5 keywords from the description and first 20 lines
• Keep the total index under 1000 tokens — truncate descriptions aggressively, not skill names
• Group similar skills mentally (e.g. all geo-* skills = "GEO/SEO domain")

Index quality

The router reads only frontmatter + first 20 lines of each SKILL.md at activation time — NOT the full content. Full SKILL.md content is only loaded when a skill is actually selected for routing. This means routing quality depends directly on the quality of each installed skill's frontmatter.

The fields that matter most:

• description: — most important. Use verb-first phrasing with concrete keywords.
• tags: — explicit keyword list for matching.
• trigger: / triggers: — alternative keyword sources.

Good frontmatter example:

description: "Run a full GEO audit on a website — crawlability, schema, content quality, AI citability"
tags: [geo, seo, audit, AI search]

Bad frontmatter example:

description: "A useful skill for various tasks"

Tip: if an expected skill is not triggering, check its frontmatter first.

Step 3 — Confirm activation

If the scan returns zero SKILL.md files, reply:

Skill Router: no skills found at ~/.claude/skills. Install skills first.

and do not activate routing.

Otherwise, reply with exactly one line:

Skill Router active [suggest | auto] — <N> skills indexed (<domain1>, <domain2>, ...).

Examples of domains: GEO/SEO, Testing, Deployment, Git, Security, Code Review, AI Delegation. List at most 5 domain labels. Nothing else.

/skill-router off

Reply with exactly one line: Skill Router off. All subsequent requests are answered directly — no routing until /skill-router on is run again.

Behavior after activation

Lifecycle note: routing state lives in conversation context. In very long sessions, context compression may truncate earlier turns and silently deactivate routing. If you notice routing has stopped, re-run /skill-router on to re-activate.

For every subsequent user input, first apply the gate check:

1. Is the request conversational, a simple question, or under ~10 words?
   → Answer directly. Do NOT load any skill.

2. Is it a substantive task (feature, audit, workflow, architecture, etc.)?
   → Scan the index. Identify all skills with HIGH confidence match.
   → Branch based on active mode (see below).

3. No HIGH confidence match?
   → Answer directly from your own knowledge. Do NOT mention any skill.

HIGH confidence = clear keyword overlap, not just tangential similarity.

✅ "run a GEO audit on plume.africa" → geo-audit (exact domain + action match)
✅ "write a Dockerfile for this Node app" → docker-patterns (explicit tool + task)
❌ "fix this bug in my app" → no match (too generic, no skill keyword overlap)
❌ "how do I structure my API?" → no match (tangential, answerable directly)

Suggest mode behavior

When one or more skills match with HIGH confidence, before answering, show:

Suggested skills for this request:
  1. <skill-name> — <one-line reason why it matches>
  2. <skill-name> — <one-line reason why it matches>  ← if a second skill also matches

Use skill 1, 2, or answer directly? (1/2/no)

• List at most 2 skills, ranked by relevance (most specific first)
• Wait for the user's reply before proceeding
• If user replies 1 or 2 → load that skill's SKILL.md fully, apply its rules, then record usage (see Usage Tracking)
• If user replies no or anything else → answer directly without loading any skill
• If only one skill matches, still show the prompt with just option 1

Auto mode behavior

When one skill matches with HIGH confidence:

• Load that skill's SKILL.md fully, apply its rules, then record usage (see Usage Tracking)
• Append exactly this footer at the end of the response:

*(via skill-name)*

• No other mention of routing. No preamble. No explanation.
• If two skills match equally, prefer the more specific one (e.g. geo-schema beats geo for a schema-only question)

Routing rules (both modes)

• Load at most one skill per request
• Never load a skill for requests under ~10 words
• Never trigger skill-router on itself
• Never auto-trigger on /skill-router status
• The bar is HIGH: only trigger when the skill would change the quality of the response, not just because keywords overlap

Usage Tracking

Every time a skill is invoked (user confirms in suggest mode, or auto-routes), do two things:

1. Update in-memory session counter:

SESSION_USAGE[skill-name] += 1

2. Persist to ~/.claude/skill-usage.json by running this Bash command (replace SKILL_NAME and TODAY):

python3 -c "
import json, os
f = os.path.expanduser('~/.claude/skill-usage.json')
d = json.load(open(f)) if os.path.exists(f) else {}
s = 'SKILL_NAME'
d.setdefault(s, {'total': 0, 'last_used': None})
d[s]['total'] += 1
d[s]['last_used'] = 'TODAY'
json.dump(d, open(f, 'w'), indent=2)
"

Where TODAY = current date in YYYY-MM-DD format ($(date +%Y-%m-%d)).

Run this silently — no output to user.

/skill-router stats

Read the persistent usage file, merge with the session counter, and display:

python3 -c "
import json, os
f = os.path.expanduser('~/.claude/skill-usage.json')
print(json.dumps(json.load(open(f)), indent=2) if os.path.exists(f) else '{}')
"

Then display a table sorted by total descending, with session column overlaid from SESSION_USAGE:

Skill Usage Stats
-----------------
Skill               | Total | Session | Last Used
--------------------+-------+---------+----------
geo-audit           |    12 |       2 | 2026-05-15
vibe                |     8 |       1 | 2026-05-18
docker-patterns     |     3 |       0 | 2026-05-01
geo-schema          |     0 |       0 | never

Rules:

• Show all skills currently in the index, even if count = 0
• Skills with total = 0 appear at the bottom with never as last used date
• Sort: descending by total, then alphabetically within ties
• Session column shows 0 for skills not used in the current session
• After the table: Total invocations this session: <sum of SESSION_USAGE values>

/skill-router status

Show exactly this, nothing more:

Skill Router status
-------------------
Mode:           suggest | auto | off
Skills indexed: <N>

Index:
  <skill-name>  |  <trigger domains>  |  <one-line description>
  ...

Last skill used: <skill-name> | none this session

If the router was never activated this session:

Skill Router not active. Type `/skill-router on` (suggest) or `/skill-router on --auto` to activate.

Non-goals

• Do NOT persist the index across sessions (stateless, session-only)
• Do NOT update the index mid-session
• Do NOT load multiple skills per request
• Do NOT trigger on greetings, status checks, or clarifying questions
• Do NOT create any files during activation or routing — except ~/.claude/skill-usage.json (usage tracking only)