alessiomarcone Navigate
Route the request. Do not absorb specialist work that belongs to a narrower skill.
Reference
Read references/knowledge-base.md when the mechanism boundary or product branch is not obvious.
Routing table
| Intent | Route |
|---|---|
| Improve or design a request | prompt |
| Fact-check or stress-test an output | verify |
| Decide, design, or configure delegated agents | subagent |
| Create or refine an agent skill | skill |
| Package or distribute reusable capabilities | plugin |
| Design a Cowork workflow | cowork |
| Operate a Codex-specific surface | codex |
| Build an official learning path | learn |
| Distill source material | knowledge-card |
| Run work on a clock or recurring schedule | cowork for Claude/Cowork, codex for Codex, or the host scheduler/API |
If no specialist matches, select the smallest supported mechanism: normal chat/task, project knowledge, CLAUDE.md or AGENTS.md, skill, subagent, hook, scheduler/automation, MCP/connector, plugin, or API/SDK. Hooks are event-driven; they are not a substitute for clock-driven scheduling.
Output contract
Return three fields, plus an optional fourth:
- Route — one skill or mechanism.
- Why — one short paragraph.
- Invoke — the exact next invocation and minimum input.
- Freshness check — only when a product capability may have changed.
Use /knowledge-pack:<skill> for the Claude Code plugin and /<skill> for a
standalone Claude skill. Use $knowledge-pack:<skill> for the Codex plugin and
$<skill> for a standalone Codex skill.
Boundaries
- Choose one adequate route, not a menu of equivalent options.
- Own comparisons such as “custom agent or skill?”; once the user explicitly selects an artifact, its specialist skill wins.
- Do not execute the specialist workflow unless routing alone would block progress.
- Do not conflate a skill, subagent, hook, plugin, connector, or MCP server.
- Answer in the user's language.