fable-prompting

Prompting Claude Fable 5

Distilled from the official guide: https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5 (Consult it for updates; this file is the working reference for prompts in this plugin.)

The one-line summary

Fable 5 is steered with brief goal-and-constraint instructions, not step-by-step scaffolding. Prompts written for Opus/Sonnet-era models are usually too prescriptive for Fable 5 and actively degrade its output. State the goal, the constraints, and the boundaries — then let it work.

How Fable prompting differs from past models

1. De-prescribe. Enumerated step lists, "CRITICAL: YOU MUST" language, and forced-cadence scaffolding ("summarize after every 3 tool calls") that older models needed now reduce quality. Prefer stating the goal and constraints over enumerating the steps. Review inherited prompts and delete instructions whose default behavior is already good.
2. One brief instruction beats an enumerated list. Instruction-following is strong enough that a single sentence steers a whole class of behavior. You do not need to name every bad pattern.
3. Give the full task specification up front. Fable 5 excels at long-horizon work when the first turn carries the complete spec — goal, constraints, definition of done. Ambiguous, drip-fed specs waste tokens and reduce quality.
4. Give the reason, not only the request. Template: I'm working on [the larger task] for [who it's for]. They need [what the output enables]. With that in mind: [request].
5. Expect longer turns. Single requests on hard tasks can run many minutes; autonomous runs for hours. Design harnesses to check in asynchronously rather than blocking, and set timeouts accordingly.
6. Delegate freely — asynchronously. Fable 5 dispatches and manages parallel subagents dependably. Don't suppress delegation (a prior-model guardrail); instead say when delegation is appropriate. Prefer long-lived subagents that keep context over spawn-and-block.
7. Fresh-context verifiers beat self-critique. For verification, spawn separate verifier subagents with clean context rather than asking the worker to critique its own output.
8. Give it a memory surface. Even a plain Markdown file. Fable 5 performs notably better when it can record and consult lessons across runs.
9. Never ask it to echo its reasoning in the response. "Show your thinking / explain your reasoning verbatim" instructions can trigger the reasoning_extraction refusal category. Read thinking blocks (API) instead, or have it report conclusions.
10. Don't surface context-budget countdowns. Remaining-token counters trigger premature wrap-up behavior. If unavoidable, add the ample-context reassurance snippet.