Use the strongest model in the world,
without paying its price every time
The pattern is called advisor and executor: Fable 5 thinks and decides, the cheaper Sonnet 5 executes. The result: most tokens at Sonnet prices, and only the big decisions on Fable.
Four rules make the pattern
Fable 5 is the advisor
It takes your task, splits it into sub-tasks, and decides which pieces need judgment (architecture calls, unclear requirements) and which are routine work.
Sonnet 5 are the executors
Routine work goes to Sonnet 5 workers: boilerplate, tests, lint fixes, and implementing any spec that is already written in detail.
Escalate, never guess
When a Sonnet worker hits ambiguity or a design decision, it never guesses. The question goes back to Fable, and Fable decides.
Isolated cache per worker
Each worker keeps its context in an isolated cache, and the full project is never re-sent. Only the delta. That is how tokens drop like crazy.
The result: Most tokens at the cheaper Sonnet price, and only the big decisions at Fable price.
The full prompt
Copy it into Claude Code and replace the bracketed part with a description of your project or task.
Set up a Claude Managed Agents architecture for this project using a two-tier delegation pattern: You (Fable 5) act as the top-level advisor/orchestrator. Break incoming tasks into sub-tasks and decide, per sub-task, whether it needs advisor-level judgment (architecture decisions, ambiguous requirements, cross-cutting tradeoffs) or can be handled by a worker. Delegate routine, well-scoped sub-tasks to Sonnet 5 sub-agents, e.g. writing boilerplate, running tests, fixing lint errors, implementing a spec you've already written out in detail. Escalate back to yourself (Fable 5) whenever a Sonnet 5 worker hits ambiguity, a design decision, or something outside its scoped instructions, rather than letting it guess. Keep each sub-agent's context cache isolated and reused. Don't re-send the full project context on every call. Structure prompts so repeat invocations of the same sub-agent only pay for new/delta context, not the full history each time. Before starting, tell me how you'd split [describe your task/project here] into advisor-level vs. worker-level pieces, and which sub-agents you'd spin up.
Before executing, Claude will show you its split: what stays with the advisor and what goes to the workers. Review the split, then let it run.
Three steps
- 1Copy the prompt and paste it at the start of a Claude Code session
- 2Replace the bracketed part with one or two sentences describing your task or project
- 3Review the split it proposes before it starts, and adjust if you want
Like the pattern?
Give your agents a foundation to build on
The same mindset is baked into Yalla Kit: a clear contract, ready-made skills, and a harness that guards quality, so your agent builds your whole Arabic product instead of guessing.