What's this about?
A concrete prompt for keeping your Fable and total usage bars in step. Pick Fable at High effort (or Medium — never Max or above), then append roughly: *"Note that your main task is analysis, orchestration and verification. Delegate concrete tasks to subagents (Opus or Sonnet) wherever possible. You should only do requirement clarification, solution decomposition, task distribution and result acceptance. Implementation work — reading large amounts of code, writing code, running tests, bulk edits — must all be dispatched to subagents via the Agent tool."* His explanation of why it works: Fable's instructions to the subagents are detailed enough that almost nothing goes wrong.
Why does the expensive model make a good manager?
Because delegation quality is a reasoning task, and it is the one place where extra intelligence pays for itself many times over. A vague instruction sends a cheap model down the wrong path for ten minutes; a precise one gets it right first time. So the expensive tokens are spent on the small amount of text that determines whether the large amount of work is useful. It is the same economics as the in-context distillation entry, arrived at independently.
Why does this matter to you?
- It is the most directly copyable thing in today's digest.
- "Never Max or above" is the specific cost warning — Max effort runs 1.5× or more.
- The two usage bars staying level is the observable signal that it is working.