Skill & Agent Router

Type /pathfinder when you need to know which skill or subagent to reach for. Nothing here fires automatically; this is a navigation aid only.

Session lifecycle

Lost a session     →  /recover-sessions   find the closed conversation and resume it
Start of session   →  /pickup
                         ↓ (do work)
End of session     →  /handoff           writes .ai/HANDOFF.md, then calls /evolve-claude-md
Standalone update  →  /evolve-claude-md  promote session knowledge to CLAUDE.md only
Periodically       →  /retro             the counterweight: propose checks, pointers, and cuts so CLAUDE.md shrinks

Skills — user-invoked (type the slash command explicitly)

Skills — model-invokable (reachable by the model or another skill mid-task)

Subagents — domain reviewer panel

Subagents are spawned with the Agent tool (subagent_type:), not slash commands. All four are review-only (report findings; never rewrite).

Review decision tree

Need a review?
│
├── Quick generalist stress-test — one skill, any domain, fast
│   → /reviewer-2
│     Baseline, counterfactual, alternatives, uncertainty consistency.
│     No domain expertise assumed.
│
├── Deep single-domain review — expert read on one dimension
│   ├── Attribution claim / result     → attribution-reviewer subagent
│   ├── Weather / synoptic mechanism   → meteo-reviewer subagent
│   ├── Statistical analysis / ML      → stats-reviewer subagent
│   └── Public-facing science product  → scicomm-reviewer subagent
│
└── Full multi-lens review — manuscript, full section, ≥2 dimensions
    → Spawn the relevant subset of the 4 subagents in a SINGLE message
      (parallel, independent — no cross-anchoring).
      Synthesize findings after all return.
      Use /reviewer-2 first if you want a fast generalist pass before the panel.

When NOT to use the panel

• Single-domain question → one reviewer, not all four.
• /reviewer-2 covers the generalist case; don't over-convene the panel for a claim check.
• Prose checks split three ways: /spot-ai = does this read as machine-written (AI markers only) · /overbaked = is it over-engineered or verbose · /reviewer-2 = is the claim defensible. spot-ai deliberately refers the other two out rather than duplicating them.
• scicomm-reviewer is for public-facing products, not internal methods or data.

Citing literature during a review

Any reviewer can flag missing citations. To retrieve them, route to /lit-review (Zotero, arxiv, bioRxiv, Google Scholar, Consensus).

Composition rule (for skills that call other skills)

A user-invoked skill may call model-invokable skills; never another user-invoked one. (Examples: handoff → evolve-claude-md; grill-me → grilling.)