Working with models and agents in a codebase¶
Read this when a task picks a model or effort level, or needs a change in code it does not own.
Model effort is set deliberately, and frugally¶
Never use the provider's default effort. A frontier model defaults to medium; a mid-tier model may default to high.
- Use the lowest effort that reliably clears the job's quality bar. [check: process]
- Set it per role, not globally. Reviews and second opinions lean higher; classification, extraction and yes/no checks run at minimum. [check: process]
- Tune from recorded real spend, never from paid calibration runs. [check: process]
Know which model is running before you start¶
- Read the actual model identity before beginning a task that was assigned to a specific model. If it differs, stop and say so before building. [check: process]
Don't dictate diffs into code you don't own¶
Send the owner of that code (agent, team or maintainer) the task and the goal, never the diff, and let it decide how. [check: process]
Check what you already have before adding manual steps¶
Before asking a person to do something by hand, check skills, the tooling reference, memory and connected servers. If a block prevents automation, say so and how to fix it; never invent a manual workaround. [check: process]