prompt-fable

Prompting Claude Fable 5

Fable 5 (claude-fable-5) is Anthropic's Mythos-class model above Opus. claude-mythos-5 is the same model without safety classifiers (limited release). This skill covers only what differs from prior Claude models; standard prompt engineering still applies.

Hard constraints

• Never instruct the model to echo, transcribe, or explain its internal reasoning in response text. This triggers stop_reason: "refusal". When migrating prompts or skills, audit for "show your thinking" / "explain your reasoning step by step in your answer" language. Any reasoning surfaced should be a concise summary.
• Safety classifiers refuse on request shape, not intent.

Deprescribe when migrating

Instruction following is strong enough that one brief instruction replaces an enumerated behaviour list. Prompts and skills written for earlier models are often too prescriptive for Fable 5 and degrade its output. When migrating: delete instructions, test default behaviour, and re-add only what measurably changes it. Heavy MUSTs, exact step sequences, and case-by-case enumeration are the smell.

Failure modes and tested mitigations

Each mitigation below is a condensed version of Anthropic's tested instruction; drop the relevant one into the system prompt when the failure appears.

• Overplans ambiguous tasks: "When you have enough information to act, act. If weighing a choice, give a recommendation, not an exhaustive survey."
• Unrequested refactoring/tidying at high effort: "Don't add features, refactor, or abstract beyond what the task requires. Do the simplest thing that works. Only validate at system boundaries."
• Fabricated progress claims on long runs: "Before reporting progress, audit each claim against a tool result from this session. If something is not yet verified, say so explicitly." (In Anthropic's testing this nearly eliminated fabricated status reports.)
• Acts when the user was only asking: "When the user is describing a problem or thinking aloud, the deliverable is your assessment. Report findings and stop; don't apply a fix until asked."
• Ends turn on a promise ("I'll now run X") deep into long sessions: "Before ending your turn, check your last paragraph. If it is a plan, a question, or a promise about work not done, do that work now with tool calls."
• Suggests wrapping up when shown a token countdown: hide context-budget counts from the model, or add: "You have ample context remaining. Do not stop, summarise, or suggest a new session on account of context limits."
• Unreadable summaries after long agentic runs (arrow chains, invented labels): "Your final summary is for a reader who didn't see the work. Outcome first, complete sentences, spell out terms, drop the working shorthand."

Scaffolding patterns that pay off

• Give the reason, not only the request. Fable 5 connects intent to context rather than inferring it: "I'm working on [larger task] for [who]. They need [what the output enables]. With that in mind: [request]."
• Subagents. Fable 5 dispatches parallel subagents readily and manages long-lived ones well. State when delegation is appropriate; prefer async communication over blocking on each return. Fresh-context verifier subagents outperform self-critique for long runs.
• Memory. Performance improves markedly given a place to record lessons across runs: one lesson per file, one-line summary at top, update rather than duplicate, delete wrong notes.
• send-to-user tool. For long async agents, a client-side tool whose input is rendered verbatim to the user. Defining it is not enough: Fable 5 rarely calls it without elicitation in the system prompt ("when you have content the user must read verbatim, call send_to_user"). Keep narration out of it.
• Aim high. Assign tasks harder than you'd give prior models; testing only simple workloads undersells the capability range.