Claude Fable 5 · Prompting · From Anthropic's own docs

How to prompt Claude Fable 5

Fable 5 follows short, clear direction better than older models. These are the few habits that change your results the most, and every one is a sentence you can paste into any chat.

The model got smart enough that piling on rules and telling it how to think backfires. Give it the why, tell it what not to touch, and make it show the receipt.
New here? Claude is an AI assistant like ChatGPT, made by a company called Anthropic, and Fable 5 is one of its newest models. Everything below works by pasting a sentence. No tools, no setup.
The shift

Smarter model, lighter touch

Older models needed long, detailed rulebooks. Fable 5 follows one clear instruction on its own, so piling on rules now gets in its way. The habits that used to help can now hurt.

Pile on a long rulebook
→
State the outcome in one line
Make it show its reasoning
→
Let it think privately, read the result
Let it run and trust "done"
→
Draw boundaries and make it prove it
The moves that matter

Six habits, six sentences

Each one is a line you paste into a normal chat, no settings or code. Fable 5 is powerful and not cheap, so a sharper prompt is also a smaller bill.

How to use these: paste the dark line into your chat and swap the [brackets] for your own words. The red line is the habit to drop. Each tip is tagged for where it works: any model or Fable-specific. Five are solid prompting anywhere, only one is Fable-specific.
1

Give it the why

Context over detail
Any model

One line of "here is why I am asking" does more than piling on step-by-step detail. Anthropic says Fable does better when it understands your intent, so the context lets it connect your task to the right information instead of guessing what you meant.

Instead of"Write me an email to a client about the delay."
Paste thisI'm working on [the bigger task] for [who it's for]. They need [what the output enables]. With that in mind: [your request].
2

Tell it what NOT to do

Negative prompting
Any model

A more capable model can act on its own more readily: drafting emails, making backups, refactoring around a one-line fix. The sentence that saves you starts with "do not."

Instead of"Take a look at this problem and handle it."
Paste thisWhen I'm describing a problem or asking a question, the deliverable is your assessment. Report what you find and stop. Don't fix, send, edit, or delete anything until I say go. Do the simplest thing that works, and skip cleanup I didn't ask for.
3

Let it act once it has enough

Stop it over-planning
Any model

More deliberation is not better. On a model that can run for minutes, endless option-surveying just burns time and money on choices it will never use.

Instead of"Research everything and make a full plan before you do anything."
Paste thisWhen you have enough information to act, act. Don't re-derive what we've already settled or narrate options you won't pursue. If you're weighing a choice, give a recommendation, not an exhaustive survey.
4

Make it prove it

Catch a fake "done"
Any model

"Done and working" is a claim, not a fact. Anthropic's own testing found this one line nearly eliminated made-up status reports, even on tasks built to provoke them.

Instead of"Is it done and working? Great, thanks."
Paste thisBefore you tell me something is done, point to the result that proves it. Only report work you can show evidence for. If something isn't verified, say so plainly instead of guessing.
5

Stop asking it to "show its reasoning"

The one that flips a rule
Fable-specific

On Fable 5, a standing "explain your reasoning" line, especially saved in a system prompt or skill, can trigger a refusal and hand your task to a less capable backup model (Opus 4.8) with no error on screen. You get a weaker answer without knowing it switched.

Delete this line"Explain your reasoning step by step, then give me the answer."
Do this insteadJust ask for the answer and let it think privately. If you need to see its thinking, read the thinking output rather than making it narrate.
6

Say less, not more

One line beats twelve
Any model

A short instruction now steers as well as spelling out most of the rules by name. Reusing last year's over-detailed prompt can actively make the answer worse.

Not a contradiction with tip 1. Add the why (that is real signal the model can't guess), and cut the rulebook (that is redundant procedure it already follows).

Instead of"Rule 1: be concise. Rule 2: use bullets. Rule 3: no jargon. Rule 4: always..."
Paste thisLead with the outcome, keep it simple, and pause only when the work truly needs me.
Tips 2 and 3 go together

They are one dial, not two rules

Tip 2 holds it back, tip 3 pushes it forward. Same model, opposite settings, and you pick which the moment needs. Match the line to the situation, and don't paste both into the same prompt.

When it's too eagerIt wants to fix, send, or refactor before you asked. Use tip 2: assess and stop.
When it's stallingIt keeps re-checking and surveying options it won't use. Use tip 3: act when it has enough.
Before you paste

Three honest caveats

These habits help a lot. They are not magic. Use them with eyes open.

Prove-it is a safeguard, not a guarantee

Anthropic's "nearly eliminated" result comes from long agentic runs where the model cites real tool results. In a plain chat with no tools, treat it as a good honesty habit, not a proven safeguard. On high-stakes work, verify it yourself, or open a fresh chat and have it check the work against your original ask.

Keep a human yes on anything irreversible

These lines cut babysitting on reversible, internal work. Keep your own approval on anything external or permanent: client emails, spending, publishing, deletions, legal, financial.

The reasoning fallback is invisible

On Claude.ai and Claude Code there is no warning when a request falls back to the backup model. You get a weaker answer from a less capable model, not a bigger bill, since the backup is actually cheaper. On the raw API you get an explicit refusal to handle instead. The fix is to remove the "show your reasoning" line.

One more thing

What happens when Fable hands off to Opus 4.8

Fable 5 runs a quick safety check on what you ask. If a request trips it, Fable steps aside and Claude Opus 4.8, Anthropic's next model down, answers instead. Here is when that happens, whether you notice, and what it costs.

You ask Fable 5Your normal request.
→
It trips a safety checkThe request looks sensitive.
→
Opus 4.8 answersA capable, cheaper backup.

1When does it happen?

When a request trips Fable's safety check: anything that looks like hacking, dangerous biology, or asking the model to reveal its own private reasoning. Ordinary coding, debugging, and security questions can occasionally get caught too.

2Do you see it?

In the Claude app and Claude Code the handoff is silent, with no on-screen note that Opus answered. If you build on the raw API, the response is clearly marked, so there you always know.

3What do you pay?

You pay for whichever model does the work, and Opus 4.8 costs half of Fable, so a handoff costs you less, not more. A request turned down before it writes anything is free.

4How do you avoid it?

Don't tell the model to show or explain its own thinking, keep security questions plainly worded, and use Opus 4.8 by default for genuine cyber or biology work.

A handoff isn't a surprise bill or a failure. You get an Opus 4.8 answer at Opus 4.8's lower price, and the only thing you give up is Fable's extra muscle on the long, hard tasks.

Source: Anthropic docs · Prompting Claude Fable 5 and Refusals and fallback

Every habit traces to Anthropic's own "Prompting Claude Fable 5" doc. Read it here.
Primary source (Anthropic):
· Docs: Prompting Claude Fable 5
· Effort · Refusals and fallback · Migrating to Fable 5
How these six were chosen: a six-lens review (aha-moment, myth-buster, cost, plain chat-box, control and trust, and retention) read the doc and argued down to the habits that change results the fastest, then each was fact-checked against the doc and Anthropic's public guidance.