Multi-model delegation in Claude Code
Why a CLAUDE.md instruction wasn't enough to make cheap models gather, mid-tier models execute, and top-tier models orchestrate - and what did work.
I work on AWS cloud infrastructure, agentic engineering, and accelerated software delivery. I write here about what I learn building it.
Why a CLAUDE.md instruction wasn't enough to make cheap models gather, mid-tier models execute, and top-tier models orchestrate - and what did work.
Boris Cherny's advice is to delete your CLAUDE.md, skills, and hooks every six months and see what the model does. These posts are what it took to answer that in a toolkit of forty-odd skills: the evals I didn't have, the ablations I ran once I had them, and what they measured.
Getting cheap models to do the gathering and expensive ones to do the deciding: why a CLAUDE.md instruction wasn't enough to make that happen, and the fan-out pattern that was.