Plans and Task Lists
The artifacts that come before code — and how to steer them with comments.
Before touching files, an Antigravity agent externalizes its intent: a task list (the checklist it will execute) and an implementation plan (files, approach, risks, test strategy). These aren't decoration — they're your cheapest steering surface. A plan that reads wrong costs you a comment; a diff that's wrong costs a rework cycle.
Read plans the way you'd read a junior engineer's design note. Did it actually look at the code first — real file paths, real function names, not generic ones? Does a test strategy exist? Is anything in scope you didn't ask for? A plan built from the task title alone is fiction, and fiction is easiest to catch before it compiles.
Steering is conversational: leave a comment directly on the plan — Google-Docs style — and the agent revises before executing. You never have to interrupt or restart to redirect.