I route Claude Code steps by decision type, not difficulty

Описание и применение

Перенаправление шагов Claude Code по типу решения, а не сложности, позволило переместить 57 % вывода на более дешевые модели, повысив продуктивность на 60 % без увеличения затрат.

Задача

Choosing the right Claude model for each code step was inefficient; hard‑task → strong‑model mapping did not improve output distribution.

Результат

Routing by decision type moved majority of output to cheaper models, increased session output by ~60%, and kept top‑model output constant, improving efficiency without higher cost.

Рабочий процесс

• 1. Identify decision type of each code step (already decided, local choice, dependent on discovery, verdict before merge).
• 2. Route step to appropriate model: Haiku for already decided, Sonnet for local choice, Opus for discovery‑dependent, Opus reviewer for verdict.
• 3. Execute step in subagent with dedicated Claude instance and context.
• 4. Return short result to main session for validation.
• 5. Main session accepts result or requests changes; merges/pushes require user approval.
• 6. Use plugin "maddog" in advisor‑mode to automate routing and validation.

Доказательства

69% of the output came from the top models (Opus, Fable)

Источник

Hard task → strong model didn’t fix it. Difficulty is the wrong signal

Источник

57% of output ran on Haiku or Sonnet, against 31% without it.

Источник

Output per session rose about 60%, while top-model output stayed flat (~124K tokens both).

Источник