How to Prompt Claude Fable5: A Practical Guide
Fable5 removed temperature and top_p, follows instructions more literally, and rewards spec-first prompting. Concrete patterns for getting the most out of the model.
Fable5 Editorial
Claude Fable5 changes the prompting contract in two ways: the classic sampling knobs are gone, and the model follows instructions more literally than any previous Claude. Both shifts reward the same habit — saying exactly what you want.
Structure every prompt the same way
Fable5 responds best to a fixed skeleton:
- Role — one line: who the model is in this task.
- Context — one line of background, not a memoir.
- Task — the actual ask, stated as an instruction.
- Output format — exact shape: headings, JSON schema, word limits.
- Hard constraints — what must never happen.
Vague prompts that older models would charitably reinterpret get executed as written. That's a feature once you adapt to it.
Spec first, not drip-fed
For long agentic tasks, put the full specification in the first message. Fable5's long-horizon focus means it builds a working plan early and keeps notes against it; corrections delivered five turns in are more disruptive than they were on Opus 4.8. Write the spec like a ticket you'd hand a senior engineer.
No temperature? Ask for variation
temperature and top_p are removed from the API. Output style is steered entirely through prompting:
- Want diversity? Say "give three structurally different approaches."
- Want conservatism? Say "prefer the boring, standard solution."
Tune effort instead
The one knob you do get is effort (low → max), which trades analytical depth against speed and cost. Practical defaults:
| Task | Effort |
|---|---|
| Classification, reformatting | low |
| Standard coding, drafting | medium |
| Architecture review, hard debugging | high |
| Frontier-difficulty problems | max |
Because output tokens cost $50 per million, unnecessary max runs are the easiest money to waste — see our cost calculator for what effort inflation does to a monthly bill.
Port your prompts down
All current Claude models share the adaptive-thinking API surface. A well-structured Fable5 prompt runs unchanged on Opus 4.8, Sonnet 4.6, and Haiku 4.5 — so prototype on a cheaper tier and escalate only where quality demands it. More on routing strategy on the homepage guide.
Want the structure done for you? The Fable5 prompt guide has copy-paste examples, and the prompt generator builds a customized system prompt in 30 seconds.