Matching each prompt to the right model, so easy work costs less.

The problem

Every prompt ran on the most expensive model, renames and git chores included. Switching by hand meant remembering to, and a fresh session meant losing the conversation.

What I did

• Wrote a hook that runs before Claude reads each prompt and scores its effort from 0 to 100 on keywords and structure. Under 20 is Haiku, up to 44 is Sonnet, up to 84 is Opus, and above that is Fable.
• Built three modes. Ask shows a picker with the recommendation first. Auto routes silently. Block rejects the prompt before the current model spends a token on it.
• Picking a different model hands the whole task to a subagent on that model, with a self-contained brief.
• Added inline tags (sm:skip, sm:haiku, sm:opus) and skips for approvals, slash commands, and short acknowledgments like "thanks".
• Taught it to ignore text the harness injects, like subagent reports and shell output, which kept opening the picker for messages nobody typed.
• Logged every decision to a file, so the routing can be audited and tuned.

Results

59 of those 66 prompts arrived in a session running Opus. Each Haiku or Sonnet recommendation was a chance to do the same work for less.

What I'd do differently

Score the work, not the wording. A keyword heuristic reads a short request that mentions git as a chore, even when it means building three pages. Next: let Haiku classify (the hook already supports it when an API key is set) and log which model I actually pick, so the thresholds learn from my overrides.

Artifacts

Source on GitHub ↗

A hook, a small config file, a README, and the decision log. The picker above is a capture of what it shows.