Model Selection Advisor Skill

The right model is rarely "the biggest one" or "the cheapest one" — it's the smallest model that clears the task's quality bar within its latency and cost budget, with a path to escalate the hard cases. This skill makes that trade-off explicit and defensible, and ties it to an eval so the choice is measured, not vibes.

Working from a brief

Given "what model should I use for summarising support tickets?", deliver a concrete recommendation anyway — infer the task's difficulty, volume, and latency sensitivity, label the assumptions, and recommend. Never hand back "it depends" with no pick; give a default and the condition under which you'd change it.

Required Inputs

Ask for these only if they aren't already provided (else infer and label):