Four documented ways to prompt Fable 5.1 that cut how fast you burn your limits

Step-by-Step Breakdown

1. Tell it what done looks like, not what to do

   Give the outcome, why it matters, what "done" means, and any real constraints — then let the model derive the task list. The doc quote behind this: Fable 5 "tends to perform better when it understands the intent behind a request. Context lets it connect the task to relevant information rather than inferring intent on its own," and this matters most "for long-running agents drawing on multiple work streams." In the demo, a long multi-clause instruction is collapsed into a single goal sentence.

2. Refactor skills written for older models

   The doc is explicit that skills built for prior models "are often too prescriptive for Claude Fable 5 and Fable 5.1." The very step-by-step skill style we were all taught to write now acts as a guard rail that slows the model down. This is the counter-intuitive one: your carefully specified skill may be costing you both quality and tokens.

3. Match effort to the task instead of leaving it on high

   The effort selector runs low / medium / high / xhigh / max, and Fable 5.1 defaults to high. The advice is to start at high, then test whether medium does the same job, then try low — against your own evals, not someone else's. Reserve xhigh and max for genuinely deep work. Two behavioural details from the docs matter here: at low effort Fable 5.1 is less likely than Fable 5 to call a search or retrieval tool and more likely to answer from memory (fine for ideation, wrong for research), and Fable 5.1 supports a per-message effort change that keeps the prompt cache, where other models require setting it at the top level of a new request.

4. Make it prove its work before it hands anything over

   Ask yourself how you would check the output if a person gave it to you — click the UI, read the sources, watch the render — and put that check in the prompt. The doc phrasings quoted: "verify your work with subagents against the specification"; "verify your work however you like"; and the strongest one, "before reporting progress, audit each claim against a tool result from the session. Only report work you can point to evidence for. If something is not yet verified, say so explicitly." The goal is to receive version five, not a first draft.

5. Parallelise and delegate so the flagship stops doing the work

   Break a task into independent pieces and have the expensive model dispatch and interpret rather than build. Herk's own practice: tell Fable 5.1 not to build, code or research, but to "spin up subagents, drive strategy, interpret what they give you, and then spin up more subagents." The docs support the mechanism — Fable 5 is "significantly more dependable at dispatching and sustaining parallel agents," it batches independent tool calls in parallel when a request names several things to fetch, and parallel subagents "save time and cost through cache reads and avoid bottlenecking on the slowest subagent." This is described as the single change that most extended his usable limit.

Commands & Code Shown

/claw-api prompt-audit

/claw-api prompt-audit

Purpose: Audits your existing skills for redundancies and rules that are no longer needed on current models — exactly the over-prescription problem the docs describe.

When to use: Run it over your skills directory once after moving to Fable 5.1. Cited in the video from a recommendation by Peter Yang; verify the command against your installed version before relying on it.

Gotchas & Caveats

• Existing Fable 5 prompts do not need rewriting to work. The docs say your Fable 5 prompts "should perform well on Claude Fable 5.1 without changes, but a handful of behavioural differences are worth knowing about." These techniques are optimisation, not migration.
• The low-effort search behaviour is a real trap. If you drop to low for cost reasons on a task that needs retrieval, you may get a confident answer from memory instead of a searched one. Effort level and tool-calling behaviour are coupled.
• Not every output has an objective check. For subjective work you are building an LLM-as-judge, which means you have to define what good looks like — the model cannot infer your taste. The claim is 97% of the way there, not 100%.
• Delegation does not move responsibility. If the parallel subagents produce a bad result, that is the orchestrating model's failure — so the verification pass still has to happen, usually at the end, with the expensive model doing it.

Key Takeaways

• Fable 5.1 defaults to high effort, and most work does not need it. Benchmarks show a wide spread across effort levels — low is roughly Fable 5 at medium/high, and cheaper.
• Per-message effort changes on Fable 5.1 preserve the prompt cache, so you can triage cheaply and solve expensively inside one conversation instead of starting a second one.
• Over-specified skills now hurt. Anthropic's docs say skills written for prior models are often too prescriptive for this one.
• Put the verification instruction in the prompt, and be specific about what evidence counts — "only report work you can point to evidence for" is the phrasing from the docs.
• Use the flagship as an orchestrator. Having Fable 5.1 dispatch and interpret rather than build is described as the biggest single saving on usage limits.
• Fable 5.1 has better out-of-the-box vision and "does its best work when it can iteratively analyse, crop, and visually verify" — relevant if you are having it check its own rendered output.