token-triage

token-triage

A Claude Code skill that decides which model, how much reasoning depth, and whether to delegate — per task — so you stop overpaying on easy work and stop redoing cheap answers on hard work.

The problem

Most people run Claude Code on one model for everything. That's the wrong default in two directions at once:

• All-Opus pays premium rates on boilerplate, formatting, and lookups a cheaper model handles identically.
• All-Haiku produces bad output on architecture and hard debugging that you then redo on Opus — paying for both runs plus your review time.

The per-token price spans ~10× (Haiku 4.5 $1/$5 → Fable 5 $10/$50 per 1M tokens). Matching the model to the task is the single highest-leverage cost decision in a session — and it's usually made by reflex, or not at all.

What token-triage does

• Recommends a model tier + reasoning depth + delegate-or-not for a task, with the reasoning shown, so you (or Claude) decide deliberately.
• Applies the same heuristic automatically when Claude spawns subagents, picking each subagent's model instead of defaulting to the priciest one.

It's grounded in real Claude Code mechanics — the Agent tool's model parameter, Explore's quick/medium/very thorough breadth, code-review's low…max effort level (the same scale as the API's output_config.effort) — not a generic complexity quiz.

Install

Copy the skill into your Claude Code skills directory:

# User-level (available in every project)
mkdir -p ~/.claude/skills
cp -r skills/token-triage ~/.claude/skills/

# — or project-level (this repo only)
mkdir -p .claude/skills
cp -r skills/token-triage .claude/skills/

Restart Claude Code (or start a new session). Invoke it with /token-triage, or let it trigger automatically when a model/effort choice comes up.

Before / after

Before: You're on Opus. You ask Claude to rename a variable across three files and summarize a changelog. Both run on Opus at $5/$25.

After: token-triage flags both as Haiku-tier mechanical work and delegates them to Haiku subagents at $1/$5 — ~5× cheaper on that slice, same result. When you next ask it to untangle a flaky async test, it keeps that on Opus at high effort, where the money is worth spending.

How it decides

The full heuristic — the model tier table and the five decision rules (cost-of-error, delegation, depth, escalation, and the Sonnet default) — lives in skills/token-triage/SKILL.md, with deeper rationale and worked examples in skills/token-triage/reference/heuristics.md.

License

MIT — see LICENSE.