Skill Catalog Orchestrator
Route complex work through the smallest useful set of skills and agents, then verify that the selected instructions actually cover the task.
Routing Workflow
- Restate the concrete outcome, risk level, and definition of done.
- Search the skill index by exact domain terms, tools, frameworks, and failure modes.
- Select the minimal skill set that covers the work; prefer specific skills over broad agent profiles.
- Read each selected SKILL.md completely before acting; load references only when the skill says they are relevant.
- Sequence skills by dependency: research, architecture, implementation, validation, review, deployment.
- Assign subagents only when independent review, parallel research, or specialized validation will improve quality.
- Resolve conflicts by following project AGENTS.md first, then user instructions, then the most specific skill, then general policy.
- After the task, note any missing skill, stale trigger, overlap, or validation gap for catalog maintenance.
Selection Rules
- Use one skill when the task is narrow.
- Use two or three skills when the task spans implementation plus validation or domain plus platform.
- Avoid loading broad agent profiles when a precise workflow skill exists.
- Avoid combining skills that duplicate the same checklist unless one adds a concrete tool or guardrail.
- Prefer official docs and active repo patterns for fast-moving integrations.
- Use review/security/compliance skills for high-impact domains even when the implementation looks small.
Catalog Health Checks
- Missing trigger: a useful skill exists but would not be invoked by the user's wording.
- Duplicate trigger: two skills describe the same action without different scope.
- Weak skill: generic advice, no checklist, no validation, no guardrails.
- Stale skill: references old APIs, deprecated tools, or outdated workflow assumptions.
- Overbroad skill: claims a domain but gives no execution path.
Output
Skill routing:
- Selected:
- Skipped:
- Order:
- Validation:
- Catalog notes: