Pick Model

Take the user's intended prompt as input. Do NOT execute it. Read the live session state, classify the prompt, emit verdict + delta + strategy: is the current model right, and if not, is switching worth the cache cost.

CLAUDE.md contract: this skill is the single source of truth for the model/effort routing call. CLAUDE.md "Execution defaults" carries no model table — it defers here by capability (auto-discovery on the description). Don't re-derive or duplicate the tier table elsewhere; if routing changes, it changes here.

Two levers, different cost (the core call):

• 🎚️ Effort change, same model → cache SURVIVES (cheap). Recommend freely.
• 🔀 Model switch → cache BREAKS, context re-read uncached (costly). Must beat the switch penalty.

A third lever — parallelism (linear vs fan-out, sub-agents vs Workflow) — is out of scope here. To design a skill/agent's per-step execution topology (and have it call this skill per step), use /pick-workflow.

Recognize-then-route: hit the right tier directly; reserve top-tier (Opus high / Fable) for ambiguous/big/can't-classify. Effort = output-spend, not input. See reference.md for principles, routing detail, escalators, examples.