model ladder

A standing rule that no delegated task silently inherits the expensive session model. Every subagent call names a model, chosen as the cheapest that will do the job well: a free local model for mechanical legs, a mid-tier for standard engineering, the top tier only for planning, judging, and synthesis.

Defaulting every task to the top-tier model burns budget with no quality gain on bounded work. Picking the rung per task keeps the expensive models for the judgment calls that actually need them.

The ladder runs bottom-up: could the free local model do this leg? If yes and I used a paid one anyway, that's a leak. Fan-outs multiply cost by the number of agents, so the discipline matters most exactly where it's most tempting to skip.

The whole method is downloadable and adaptable — method/steps/model-ladder.md inside the zip. The local rung is any capable open-weight coder served on your own machine; the practice transfers to whatever tiers you have.