The two official patterns from Anthropic, plus the exact setup I used on camera. Benchmarks are Anthropic's own: advisor pattern scored 92% of Fable 5's quality at 63% of the price (SWE-bench Pro), orchestrator scored 96% at 46% (BrowseComp).
The idea in one line: stop letting the expensive model do the typing. Put it where judgment matters, let cheap models do the volume.
Your everyday session runs on a cheap model (Sonnet). It calls Fable only when it hits something genuinely hard, about once per task.
From your project root, create the file .claude/agents/fable-advisor.md with this content:
---
name: fable-advisor
description: Senior advisor for hard problems. Consult when stuck on architecture decisions, ambiguous tradeoffs, or when an approach isn't working. Returns guidance, not implementation.
model: claude-fable-5
---
You are a senior technical advisor running on Fable 5. The main session
does the work. Your job is judgment, not typing.
When consulted: analyze the problem, identify what the executor is missing,
and return a short, decisive recommendation with reasoning. Do not write
the implementation yourself. Keep answers under 300 words so escalations
stay cheap.
Then in Claude Code:
- Restart the session (agent files load at start), or create the agent via /agents > Create new agent
- Check /agents > Library shows fable-advisor
- Switch your session to the cheap model: /model sonnet
- Give it a task and tell it the advisor exists:
You're the executor on this task: [your task]. When you hit an architectural
decision you're not confident about, consult the fable-advisor sub-agent for
guidance, then implement its recommendation yourself.
Real numbers from my run: whole task $0.37, the Fable consultation was $0.14 of it, one call, exactly as the pattern describes.
If Fable's model string differs on your plan, check the exact name with /model and update the agent file's frontmatter.
Flip it: Fable makes the plan and delegates the token-heavy work to cheaper Sonnet workers. The expensive model is the manager, and managers don't type.
Anthropic published the exact notebook: https://github.com/anthropics/claude-cookbooks/blob/main/managed_agents/CMA_plan_big_execute_small.ipynb
And the advisor tool docs: https://platform.claude.com/docs/en/agents-and-tools/tool-use/advisor-tool
From a Claude Code team member's field guide: Fable is the first model where the quality of the work is bottlenecked by how well you explain what you want. The model stopped being the limit.
Before running either pattern, run this once:
Interview me, one question at a time, about anything ambiguous in this task.
Prioritize questions where my answer would change the architecture.
Two more from the same guide worth stealing:
Blindspot pass (before unfamiliar work):
I'm working on [thing] but I don't know this area well. Do a blindspot pass:
help me find my unknown unknowns so I can prompt you better.
The quiz (after big changes, before you merge):
Give me a report on everything that changed, with context and intuition,
and a quiz at the bottom that I must pass before we merge.
The model leaves subscriptions Sunday July 12. The patterns work regardless, they're how you run any expensive model without paying expensive-model prices.
Built by @kju4q, more at github.com/kju4q