AI Catchup

Prompting Claude Fable 5: What to Change and What to Delete

By 5 min read

Prompt Claude Fable 5 with less, not more: brief instructions now beat enumerated rule lists, and old skills written for prior models can degrade output. Use effort as your main cost control, expect longer turns, ground progress claims in tool results, and configure fallback to Opus 4.8 for safeguard refusals.

Claude Fable 5 is the model where the prompting advice flips: instead of adding instructions, you delete them. Anthropic's own guidance says skills and prompts written for prior models are often too prescriptive for Fable 5 and can degrade output, and that a brief instruction now does the work of an enumerated behavior list. This guide distills Anthropic's official Fable 5 prompting guide into the changes worth making, in priority order.

This is the model-specific layer. For techniques that apply to every current model, start with our general prompt engineering playbook; for the other frontier models, see prompting Claude Opus 5 and prompting GPT-5.6.

Key Takeaways

  • Give it your hardest problems. Anthropic says teams see the best outcomes applying Fable 5 to their hardest unsolved problems; testing it only on simple workloads undersells its range.
  • Effort is the primary control. Default to high, use xhigh for capability-sensitive work, and drop to medium or low for routine tasks. Lower effort on Fable 5 often exceeds xhigh on prior models.
  • Shorten your prompts. Instruction following is strong enough that a one-line steer replaces an enumerated list. Audit old skills and remove what the model now does by default.
  • Plan for longer turns. Hard tasks can run many minutes per request and hours per autonomous run. Adjust client timeouts and check on runs asynchronously instead of blocking.
  • Ground progress claims. Tell the model to audit every status claim against a tool result; Anthropic says this nearly eliminated fabricated reports in testing.
  • Configure safeguard fallback. Safety classifiers can refuse offensive-security, life-sciences, and reasoning-extraction requests. Route refusals to Claude Opus 4.8 automatically.
  • Give it a memory. Fable 5 performs notably better when it can write lessons to files and reference them across runs.

Start at the Top of Your Difficulty Range

Fable 5 is built for end-to-end work that takes a person hours, days, or weeks: multiday goal-directed runs, single-pass implementations of systems that previously took days of iteration, and ambiguous multithreaded requests where the model determines next steps itself. Anthropic's explicit scaffolding advice is to pick a task harder than what you would assign to prior models, then have Fable 5 scope it, ask clarifying questions, and execute.

The practical consequence is turn length. Individual requests at higher effort can run for many minutes, and autonomous runs can extend for hours. Before migrating an agent harness, raise client timeouts, add streaming and progress indicators, and restructure blocking loops to check on runs asynchronously. If the model overplans on ambiguous tasks, Anthropic's sample steer is direct: when it has enough information to act, act, and give a recommendation rather than an exhaustive survey.

Delete Instructions Before You Add Them

Instruction following improved enough that most behaviors take a one-line steer. Instead of listing every verbosity pattern you want suppressed, one brevity instruction ("lead with the outcome; supporting detail after") covers surveying options, long root-cause explanations, and over-structured output. The same applies to checkpoint behavior: "pause only for destructive actions, real scope changes, or input only the user can provide" replaces a page of cases.

Audit existing prompts and skills when you migrate. Anthropic warns that prescriptive skills tuned for prior models can actively degrade Fable 5's output, and that default performance is often better with the old instructions removed. One instruction class is now actively harmful: anything telling the model to echo, transcribe, or explain its internal reasoning in the response can trigger the reasoning_extraction refusal category and elevate fallbacks. If you need reasoning visibility, read the structured thinking blocks from adaptive thinking instead.

Ground Progress Claims During Long Runs

On long autonomous runs, instruct Fable 5 to audit progress against actual tool results before reporting: only report work it can point to evidence for, flag anything unverified, and state failures plainly with output. Anthropic says this pattern nearly eliminated fabricated status reports in its testing, including on tasks designed to elicit them. Pair it with explicit boundaries, because Fable 5 can occasionally take unrequested actions (drafting an email nobody asked for, creating defensive git-branch backups): when the user is describing a problem rather than requesting a change, the deliverable is an assessment, not a fix.

Use Subagents and a Memory System

Two scaffolding investments pay off disproportionately on Fable 5:

InvestmentWhy it works on Fable 5
Parallel subagentsAnthropic says the model is significantly more dependable at dispatching and sustaining parallel subagents; prefer async communication over blocking on each return
Fresh-context verifier subagentsSeparate verifier agents tend to outperform self-critique on long-running builds
A memory file systemThe model performs particularly well recording lessons from previous runs, one lesson per file, and referencing them later
A send-to-user toolLong async agents need a way to surface verbatim content mid-turn; define the tool and prompt for it, because the model rarely calls it unprompted

The memory system can be as simple as a Markdown directory. Anthropic's guidance: store corrections and confirmed approaches with the why, skip anything the repo already records, update rather than duplicate, and delete notes that turn out wrong.

Know the Safeguards and the Two Rare Failure Modes

Fable 5 runs safety classifiers targeting offensive cybersecurity techniques, biology and life sciences content, and extraction of its summarized thinking, and benign work in those areas can trigger them. Requests can return stop_reason: "refusal"; configure server-side or client-side fallback to Claude Opus 4.8 so declined requests re-route automatically. Also note the API differences: adaptive thinking only, summarized-only thinking output, and no extended thinking budgets.

Two behaviors show up rarely but are worth pre-empting in system prompts. Deep into long sessions the model can end a turn with a statement of intent ("I'll now run X") without the tool call; an autonomy reminder ("you are operating autonomously; do the work now rather than ending on a promise") addresses it. And in very long sessions it can suggest wrapping up out of context-budget concern, most often when the harness shows it a remaining-token countdown; hide the countdown where possible, or add a reassurance that context is ample.

Finally, give the reason, not only the request. Fable 5 performs better when it knows why you are asking: "I'm working on X for Y; they need Z; with that in mind: [request]" beats the bare request, especially for long-running agents juggling multiple workstreams.

Sources

Read next

Keep building the workspace playbook

Frequently Asked Questions

What effort level should I use with Claude Fable 5?

Anthropic recommends high as the default, xhigh for the most capability-sensitive workloads, and medium or low for routine work. Lower effort settings still perform well and often exceed xhigh performance on prior models, so re-run an effort sweep instead of carrying over old defaults.

Do my existing prompts and skills work on Fable 5?

Mostly, but Anthropic warns that skills developed for prior models are often too prescriptive for Fable 5 and can degrade output quality. Review old instruction lists and remove what the model now does by default; a brief instruction usually replaces an enumerated list of behaviors.

Why is Fable 5 refusing my request?

Fable 5 runs safety classifiers targeting offensive cybersecurity, biology and life sciences content, and extraction of its summarized thinking. Benign work in those areas can also trigger them, returning a refusal stop reason. Anthropic recommends configuring automatic fallback to Claude Opus 4.8 for declined requests.

How do I stop Fable 5 from fabricating progress reports?

Instruct it to audit each claim against a tool result from the session before reporting, and to explicitly flag anything not yet verified. Anthropic says this nearly eliminated fabricated status reports in its testing, even on tasks designed to elicit them.

Get the weekly AI Catchup

Tools, practices, and what matters, in your inbox every week.