← All guides
Business · Tools ·Claude

Claude Fable 5: What It Is and How to Prompt It

Fable 5 is Claude's most capable model, built for big, hours-long jobs it can run start to finish. The twist: you prompt it less, not more. Give it the goal and the reason, set the boundaries, and let it work. Here is the exact approach, plus a copy-paste starter block.

Most guides about a new AI model are a spec sheet. This one is not, because Fable 5 changes something more useful than the numbers: it changes how you should talk to it.

Fable 5 is Anthropic’s most capable widely released model. It is built for the hard, long, ambiguous jobs that used to be too big to hand an AI: the kind of work that takes a person hours, days, or a full week. Early testers reported it doing in one pass what used to take days of back-and-forth. That is the headline.

But here is the part nobody tells you, and the reason I wanted this guide up first. With Fable 5, you prompt less, not more. The heavy, spell-out-every-step prompting that makes weaker models behave will actually make Fable 5 worse. It is strong enough that over-coaching gets in its way. So the skill is not writing a longer prompt. It is handing off the work like you would to a sharp new teammate, then getting out of the way.

Let me show you what that looks like.

First, what Fable 5 is actually for

This is not the model for a quick email or a fast summary. It is a premium tool, priced and built for the big jobs. Use a standard model for the everyday stuff (here is how I pick between the tools), and save Fable 5 for the work that has real weight.

Where it pulls ahead of everything before it:

  • Long-horizon work. It stays on task and holds the thread across long, multi-step jobs without wandering off or forgetting the goal.
  • First-shot correctness. Give it a well-specified hard problem and it often nails it in one pass instead of five rounds of fixes.
  • Real deliverables. Financial analysis, spreadsheets, slide decks, documents. It stays in scope and produces professional-grade output, not a rough draft you have to rebuild.
  • Ambiguity. Hand it a messy, multi-threaded request and ask it to figure out the next steps, and it does.

The teams getting the most out of it are the ones pointing it at their hardest unsolved problems, not testing it on easy ones. If you only give it simple work, you will never see what it can do.

The mindset shift: brief it, don’t script it

If you have read my 5-part prompt guide, you know the structure: role, task, context, format, example. That structure still helps you start. But with Fable 5, the weight shifts hard toward two of those parts and away from the rest.

The two that matter most:

  1. Task, stated as a goal, not a checklist. Tell it the outcome you want, not the ten steps to get there. It will find the steps.
  2. Context, especially the why. This is the single biggest lever with Fable 5, and it is the one people skip.

Fable 5 does noticeably better work when it understands why you are asking. The reason lets it connect the task to what actually matters instead of guessing at your intent. So do not just give it the request. Give it the reason behind the request.

Here is the template I use, and it is almost embarrassingly simple:

I'm working on [the larger goal] for [who it's for].
They need [what this output lets them do].
With that in mind: [your actual request].

Compare the two:

  • Just the request: “Build a mockup of this new product idea.”
  • The request plus the reason: “I’m exploring whether we should build [the feature] into our product. Leadership needs to decide if it’s worth a full quarter of engineering time, so they need to see it working and understand exactly how we’d bring it to market. With that in mind: build a working mockup of the feature and the go-to-market plan for launching it.”

Same starting ask. The second one gets you a clickable mockup built around an actual launch decision, not a rough sketch you have to explain in a meeting. You did not write a longer prompt. You told it what the work is for.

Set the boundaries, then let it run

Because Fable 5 is so capable and so willing to work, the second half of prompting it well is telling it where the edges are. A strong model left unbounded can over-deliver: adding features you did not ask for, tidying things that were fine, drafting an email when you only wanted an opinion. The fix is not to micromanage the steps. It is to name the boundaries up front.

These are the boundary lines I paste into big jobs. They are plain-language versions of what Anthropic itself recommends, and they matter more the harder you push the model:

  • Act when you have enough. “When you have enough information to act, act. Don’t re-ask me things we’ve already settled, and don’t list options you won’t actually pursue.”
  • Keep it simple. “Do the simplest thing that works. Don’t add features, cleanup, or ‘just in case’ handling I didn’t ask for.”
  • Report honestly. “Before you tell me something is done, check it against what you actually produced. If something isn’t verified yet, say so plainly.”
  • Stop only when it matters. “Only pause to ask me when the work truly needs it: a decision only I can make, or something you can’t undo. Otherwise keep going.”
  • When you’re thinking out loud, that’s fine. When you report to me, lead with the outcome. “End with one plain sentence on what happened or what you found, then the details.”

That last one is worth calling out. Deep into a long job, Fable 5 can start writing to you in its own shorthand: dense, technical, full of references to work you never saw it do. Telling it to lead with the outcome and write the final summary for someone who wasn’t watching fixes it. You get a report you can actually read.

The copy-paste starter block

Put the goal, the reason, and the boundaries together and you get a starter block you can paste at the top of any big Fable 5 job. Fill in the brackets and go:

Here's the job: [what you want done, start to finish].
Why it matters: I'm doing this for [who], and they need it so they can [what the output enables].

As you work:
- When you have enough to act, act. Don't re-ask settled questions or list options you won't use.
- Do the simplest thing that works. No extra features, cleanup, or "just in case" handling I didn't ask for.
- Before you say something is done, check it against what you actually produced. If it isn't verified, say so.
- Only stop to ask me when it's truly necessary: a decision only I can make, or something you can't undo.
- When you finish, lead with the outcome in one plain sentence, then the supporting detail.

Notice there is not a single “step 1, step 2, step 3” in there. That is on purpose. You told it the outcome, the reason, and the rules of the road. The steps are its job.

A few things that will feel different

Fable 5 does not behave like the chat models you are used to, and knowing that ahead of time saves you a lot of “is it broken?” moments.

It takes longer, and that is normal. A single request on a hard task can run for many minutes while it works and checks itself. Do not interrupt it. If anything, the mistake people make is pulling it off a job too early. Let it finish.

It can decline some requests. Fable 5 has safety classifiers that can turn down certain topics, mostly offensive cybersecurity and biology or life-sciences work. It declines cleanly rather than throwing an error, and occasionally legitimate work near those areas gets caught too. If you are building something serious on it, the standard move is to have it automatically hand declined requests to another Claude model so the job still gets done.

Give it a memory for repeat work. Like any Claude model, Fable 5 gets better when you let it keep notes between sessions. If you are running the same kind of job over and over, give it a memory so it carries forward what worked and what didn’t instead of starting cold every time.

Don’t tell it to “show its thinking” in the answer. This one is specific to Fable 5. Asking it to transcribe or explain its internal reasoning inside the response can trip a safeguard and get the request bounced to another model. If you want to see how it got there, use the model’s own summary view for that, and keep your prompt focused on the work.

Do this today

Pick the idea you have written off as too big to try. Not a quick task, and not your most annoying chore, the thing you assumed needed a team, a budget, or a specialist you do not have, so it never left the notes app. A product mockup. A full agentic workflow to run part of your business. A tool you would have hired someone to build. Write it up with the starter block above: the goal, the reason it matters, and the five boundary lines. Then hand it to Fable 5 and let it run.

The shift you are practicing is the one that matters most as these models get stronger: you are learning to delegate to AI, not dictate to it. Brief it like a capable teammate, tell it what the work is for, name the edges, and trust it with the middle. That is the whole skill.

New to all of this? Start with your first week with AI and come back when you have a job big enough to be worth this model.

Want my full Fable 5 starter prompt, the exact boundary lines I reuse, and the fill-in-the-blank brief I use to hand it work like a teammate? Grab them below and I will send them straight over.

Common questions

Is Fable 5 the model I should use for everyday chat?

No, and this is the part people get wrong. Fable 5 is a premium power tool built for big, long, hard jobs it can run end to end. For a quick email, a summary, or a fast question, a standard model is faster and cheaper and just as good. Reach for Fable 5 when the job would take a person hours or days, not seconds.

Where do I actually get Fable 5?

It is Anthropic's flagship, available through the Claude API and the major cloud platforms (AWS, Google Cloud, Microsoft Foundry). In plain terms, it is the model builders reach for. If you use an AI agent or an automation built on Claude, you may already be running it under the hood. It is not the model you casually pick for a one-line chat.

Why does it sometimes take minutes to answer?

Because that is the point. Fable 5 is built for long-horizon work, so a single request on a hard task can run for many minutes while it gathers context, does the work, and checks itself. Autonomous runs can go for hours. Longer waits are normal here, not a sign something is stuck. Let it run.

Why did it decline my request?

Fable 5 has built-in safety classifiers that can decline certain requests, mostly around offensive cybersecurity and biology or life-sciences topics. When it declines, it does so cleanly rather than erroring. Legitimate work near those topics can occasionally trip it too. Builders usually set it to automatically hand those requests to another Claude model so the work still gets done.

Do my usual 5-part prompt rules still apply?

Mostly, but dial the scaffolding back. The task and the context (especially the why) matter more than ever. The rigid step-by-step scripting matters less, and can actually make Fable 5 worse. Give it the goal, the reason, and the boundaries, then trust it to figure out the steps. That is close to the opposite of how you coach a weaker model.

The Newsletter

One practical AI idea, every week.

Real ways I use AI at work and at home. No hype, no jargon, no tech background needed. Coming soon for subscribers: the free course, 5 Days to Actually Using AI.

Free forever. Unsubscribe anytime.