What's this about?
Two things. A repository of Fable 5.1 world-modeling experiments. And a `fable-advisor` plugin for Claude Code that runs a three-stage loop: Fable 5.1 orchestrates, GPT-5.6 Luna implements, Fable 5.1 reviews. The author's framing: think of it as in-context distillation — giving Luna the intelligence of Fable 5.1 at a fraction of the cost.
What is "in-context distillation"?
Ordinary distillation trains a small model on a large one's outputs, permanently. This does it per-task and at runtime: the expensive model writes the instructions, so the cheap model receives a much better-specified problem than a human would have given it, and the expensive model checks the result. The capability transfer happens in the prompt rather than in the weights. dotey's explanation of why it works is the useful one — Fable's instructions to subagents are detailed enough that very little goes wrong.
Why is it worth noticing?
- Cross-lab orchestration is a genuinely new pattern and today produced two independent versions of it — this one and dotey's Fable-directs-Opus prompt.
- It is a direct economic response to the week's quota complaints: use the expensive intelligence for judgement, not for typing.
- dotey's own take: he does not bother, because having Fable direct Opus is simpler; the cross-lab version makes more sense on API billing than on a subscription.