Research. Experiment. Adopt.
154 AI finds from Sandy's Instagram bookmarks, turned into experiments anyone can run. Try one, check off its steps, level up.
A usage-limit-stretching loop for Claude Code: the top model (Fable 5 or Opus) is too powerful to waste on tiny tasks, so it acts as planner and judge while an executor subagent on a cheaper, faster model (Sonnet or Haiku) does the grunt work, and the top model reviews the final result. The /goal command (a real built-in, v2.1.139+) keeps the work running until it hits stated criteria, and the video suggests a loop or schedule for full autonomy. Same advisor/tiering pattern as the Fable 5 workshop row (row 80).
Set up a model-tiering loop in Claude Code for a task I will describe. Configure an executor subagent that runs on a cheaper, faster model (Sonnet or Haiku) for the grunt work, while this session's top model writes the plan first and reviews the executor's output at the end, flagging anything that needs rework. Then propose a /goal completion condition for the task phrased as objectively verifiable criteria (tests pass, files under a size budget, checklist items done) so the loop continues until the work is actually finished. Keep all runs supervised in-session; do not set up any scheduled or unattended execution. The task is: