Skill router

Skill Router

Use this skill as a default routing layer before substantial Codex work. Keep routing silent for simple low-risk tasks. Show a routing summary only when the task is complex, ambiguous, multi-skill, or high-risk.

Core Decisions

Decide these before acting:

• Relevant skills to use.
• Goal mode or Plan mode.
• Whether multiple skills should be chained.
• Whether read-only inspection must happen first.
• Whether the task contains high-risk operations.

Silent vs Visible Routing

Use silent routing when the task is simple, low-risk, and clearly scoped, such as small edits, explanations, one-file bugs, or direct questions.

Show a short routing summary when any of these apply:

• New project, architecture, scaffolding, or multi-file implementation.
• Browser automation, login, account/session workflow, or screenshots.
• Database, Supabase, migration, auth, RLS, production data, or schema changes.
• GitHub PR, issue, CI, push, release, publish, or deployment.
• Config, environment variables, secrets, credentials, or system settings.
• Deleting files, moving many files, cleanup, reset, overwrite, or destructive operations.
• Multiple skills are likely needed.
• The user explicitly asks for planning, safety boundaries, or skill selection.

Mode Selection

Recommend Goal mode for low-risk work:

• Small clear fixes.
• Single-file or narrow changes.
• Clear bug with direct reproduction.
• No account, browser, database, config, GitHub write, deletion, or release risk.

Recommend Plan mode for higher-risk or broader work:

• New project or multi-file design.
• Architecture decisions.
• Browser automation, login, or account actions.
• Database work, migrations, Supabase, RLS, auth, or production data.
• Config changes, environment variables, secrets, credentials, or tool setup.
• GitHub operations, CI debugging, PR review, push, release, or deployment.
• File deletion, cleanup, reset, bulk moves, or overwrite operations.

If a task includes deletion, config/env changes, login, browser automation, database work, GitHub push/release, or deployment, recommend Plan mode and start with read-only inspection.

Complex/high-risk tasks must stop after plan/read-only inspection and wait for explicit user approval before write actions.

Skill Matching

Prefer discovered skills from data/skills_cache.json when available. Use data/custom_rules.json to improve unknown or local skills. If the cache is missing or stale, ask to run the refresh script from the installed skill directory:

python skill-router/scripts/refresh_skills.py

Use common pairings:

• PDF to slides: pdf then presentations.
• Spreadsheet or CSV analysis: spreadsheets and, for analytical decisions or dashboards, data-analytics.
• GitHub PR or CI: github, then gh-fix-ci or gh-address-comments when relevant.
• OpenAI API app: openai-platform-api-key, then openai-docs or OpenAI developer skills.
• Supabase or Postgres: supabase, then supabase-postgres-best-practices for SQL/schema/performance.
• Browser testing or local web UI: browser or playwright-interactive.
• Desktop app automation: computer-use or electron when applicable.

If a discovered skill has no registry/custom explanation, mark it as unknown, keep it visible as a candidate, and suggest adding metadata to data/custom_rules.json.

Visible Output Template

For complex or high-risk work, keep the summary short:

Routing:
- Mode: Plan
- Skills: skill-a -> skill-b
- Read-only first: yes
- Risk flags: config change, browser automation
- Boundary: inspect first; do not modify/delete/push until plan is confirmed

Then continue according to the active collaboration mode and user instructions.