Why Fable prompts differently
Fable 5 isn't a faster Opus — it's built for a different shape of work. Anthropic's guide says the teams getting the most out of it assign it their hardest unsolved problems; testing it on simple workloads undersells it. The behavioral shifts that matter:
- It runs long. Single requests on hard tasks can run many minutes; autonomous runs extend for hours or days, with strong instruction retention throughout.
- It gets things right the first time more often. Early testers reported single-pass implementations of systems that previously took days of iteration.
- It manages fleets. Fable dispatches parallel subagents readily and dependably — including long-lived ones that keep their context across subtasks.
- It sees better. Dense screenshots, charts and technical images read substantially more accurately — it will even crop and clean up bad images itself.
- It reviews code unusually well. Bug-finding recall is noticeably higher than Opus 4.8, including searches across whole codebases and repo history.
One correction to a claim making the rounds: Fable 5 is not "also known as Mythos." They're siblings — the same underlying model, but Mythos 5 ships without Fable's extra dual-use safety measures, to approved organizations only.
1 · Pick the effort level like you'd staff a job
Effort is the primary dial trading intelligence against latency and cost. The guide's advice, translated to a hiring decision:
low/medium— quick questions, rewrites, routine work. Still strong; often beats maximum effort on older models.high— the default.xhigh— the problems where quality is non-negotiable.
Two symmetric warnings: drop the effort if routine tasks are taking longer than they should (Fable will happily over-deliberate a simple job), and raise it when you want its best verification and reasoning behavior.
2 · Give the reason, not only the request
Fable performs best when it knows why you're asking — context lets it connect the task to everything else it can see instead of guessing your intent. Anthropic's template:
I'm working on [the larger task] for [who it's for].
They need [what the output enables].
With that in mind: [your request].3 · Short instructions beat long lists
This is the counterintuitive one. Instruction-following is strong enough that one brief steering sentence replaces the page of rules you wrote for older models — and over-specified prompts can actually degrade output, because you're constraining a model that would have found a better approach on its own. Anthropic's own brevity instruction is two sentences:
Lead with the outcome. Your first sentence after finishing should answer
"what happened" or "what did you find." Supporting detail comes after.
Keep output short by being selective about what you include — not by
compressing the writing into fragments or jargon.Corollary from the migration notes: skills and saved prompts built for Opus-era models are often too prescriptive for Fable and can produce worse results than a fresh, simple prompt. Re-test them; delete what no longer earns its keep.
4 · Set the checkpoints, or it will set its own
Fable is built to keep going. For sensitive work, say exactly when it should stop and ask:
Pause for the user only when the work genuinely requires them: a
destructive or irreversible action, a real scope change, or input that
only they can provide. If you hit one of these, ask and end the turn,
rather than ending on a promise.And its mirror image, for pipelines meant to run unattended: tell it plainly that nobody is watching, that reversible actions should proceed without asking, and that it should end its turn only when the work is done or truly blocked.
5 · For long runs: memory, honesty, boundaries
Memory. Fable improves markedly when it can write lessons down and read them back later. A folder of Markdown files is enough. Anthropic's recommended instruction:
Store one lesson per file with a one-line summary at the top. Record
corrections and confirmed approaches alike, including why they mattered.
Don't save what the repo or chat history already records; update an
existing note rather than creating a duplicate; delete notes that turn
out to be wrong.Honest status reports. On long autonomous runs, make it audit its claims against actual tool results — in Anthropic's testing this nearly eliminated fabricated progress reports:
Before reporting progress, audit each claim against a tool result from
this session. Only report work you can point to evidence for; if
something is not yet verified, say so explicitly.Boundaries. Fable can occasionally do more than you asked (drafting the email nobody requested, making "just in case" backups). One instruction fixes the pattern: when the user is describing a problem or thinking out loud, the deliverable is an assessment — report findings and stop; don't fix until asked.
If it stalls. Deep into a long session it may rarely announce a step without taking it, or ask permission it doesn't need. A plain "go ahead and do it end to end" restarts it.
6 · Verification: fresh eyes beat self-review
The guide's scaffolding advice worth stealing even for casual use: separate, fresh-context verifier subagents outperform self-critique. For long builds, tell Fable to establish a method for checking its own work at an interval and to run that check with subagents against the spec — don't just trust the builder's own report.
The caveats, honestly
- It costs real money. $10/$50 per MTok — 2× Opus 4.8. On claude.ai plans it's included only through July 7, 2026 (at 50% of weekly limits), then it needs usage credits. Current terms →
- Safety classifiers are part of the deal. Since the July 1 return, classifiers watch offensive-cybersecurity and biology/life-sciences territory — and benign work in those areas can trip them. Declined requests can fall back to Opus 4.8 automatically. The trade-offs →
- Don't ask it to transcribe its own reasoning. Prompts that say "show your thinking in the response" now trigger a refusal category and force fallbacks. Read the structured thinking output instead.
- Longer turns are normal. Budget for minutes, not seconds, on hard tasks — and check on long runs asynchronously instead of staring at a spinner.
Moral: brief it like a senior hire — the goal, the reason, the boundaries — then get out of the way.
Sources
Primary: Prompting Claude Fable 5 — Anthropic docs (all quoted instruction blocks are Anthropic's recommended prompts, reproduced for use). Companion reading: Introducing Claude Fable 5 and Claude Mythos 5. A community plain-English pass by the AI Edge newsletter circulated July 2026 and prompted this recipe; where it and the primary source disagree (the "Fable = Mythos" naming, the expired June 22 window), this page follows the primary source.