Premium Router — Development skill for Claude Code

premium-router

**Fable brain, Haiku price.** A Claude Code Skill that gives you Claude Fable 5 quality while paying close to Claude Haiku 4.5 rates.

One principle drives everything here: **Fable does the judgment, Haiku does the volume.** Writing good decision rules is hard (Fable's job, paid once). Applying them is easy (Haiku's job, paid at 1/10th the rate). Reading a mountain of documents is volume (Haiku workers). Deciding what the mountain means is judgment (Fable, reading only the compressed briefs).

Everything in this repo was measured live on the API, including the approaches that FAILED. The anti-patterns section is as important as the features.

Quickstart (5 minutes)

# 1. Install
git clone https://github.com/ToolMonsters/premium-router ~/.claude/skills/premium-router
cd ~/.claude/skills/premium-router && pip install -r requirements.txt
export ANTHROPIC_API_KEY="your-key"   # never committed

# 2. Teach it once: Fable writes the rulebook from your labeled examples
python route.py distill --task-type my_task --examples my_examples.txt

# 3. Use it forever: Haiku applies the rulebook, ~10x cheaper
python route.py classify --task-type my_task --input "a new message"

# 4. For document analysis: Fable judges, Haiku workers read when it is big
python route.py analyze --mission "Prioritize these accounts for this week" \
  --docs thread1.txt thread2.txt transcript3.txt

The repo ships with a ready-made rulebook (`sales_reply_classification`), so you can skip step 2 to try it immediately.

The three commands

Command	Who thinks	Who does volume	When to use
distill	Fable (effort high, once)	nobody	teach a new repeatable task
classify	Haiku (Fable on escalation)	Haiku	every repeatable request
analyze	Fable (effort low)	Haiku workers if corpus is large	judgment over documents

Measured results (live on the API, not estimates)

classify: same accuracy as Fable, ~10x cheaper