Claude Fable 5 prompting.

To prompt Claude Fable 5 well, describe the outcome you want and the reason you want it, set clear boundaries, and let the model decide the steps. That is the core shift Anthropic describes in its official Fable 5 prompting guide: instruction following is now strong enough that one short, plain instruction replaces the long lists of rules older models needed. This guide turns that documentation into a step-by-step recipe a non-technical reader can use today.

Claude Fable 5 became generally available on June 9, 2026. It reads up to 1 million tokens of context (roughly 750,000 words), writes up to 128,000 tokens of output in one response, and costs $10 per million input tokens and $50 per million output tokens on the Claude API, per Anthropic's pricing documentation. If you use Claude through the app rather than the API, the prompting advice below applies the same way.

• Claude Fable 5 reached general availability on June 9, 2026 with a 1 million token context window and up to 128,000 output tokens per response (Anthropic documentation).
• API pricing is $10 per million input tokens and $50 per million output tokens, with a 50% discount for batch processing (Anthropic pricing documentation).
• Anthropic's guidance: one short instruction now steers behavior that previously required enumerating each rule by name (Anthropic, Prompting Claude Fable 5).
• The effort setting has four levels (low, medium, high, xhigh); Anthropic recommends high as the default for most tasks.
• In Anthropic's testing, asking the model to audit progress claims against actual tool results nearly eliminated fabricated status reports on long runs.
• Anthropic recommends starting at the top of your difficulty range: teams that test Fable 5 only on simple tasks undersell what it can do.

What is Claude Fable 5, and what actually changed?

Claude Fable 5 is Anthropic's most capable model, built for long, complex work: tasks that take a person hours, days, or weeks. Compared with the previous flagship (Claude Opus 4.8), Anthropic's documentation lists improvements in long-horizon autonomy, first-shot correctness on well-specified problems, reading dense screenshots and documents, enterprise output like financial analysis and slides, code review, handling ambiguous requests, and managing teams of parallel sub-agents.

For you, the practical change is this: the model needs less supervision per step and more clarity up front. Older models rewarded prompts that spelled out every step. Fable 5 rewards prompts that spell out the destination, the purpose, and the fences, then get out of the way. Anthropic puts it directly: skills and prompts written for prior models are often too prescriptive for Fable 5 and can degrade output quality.

One more change worth knowing before you prompt anything: turns are longer. A hard request can run for many minutes while the model gathers context, builds, and checks its own work, and autonomous runs can extend for hours. That is normal behavior, not a hang. Anthropic explicitly advises adjusting timeouts and expectations before assuming something is wrong.

How do you prompt Claude Fable 5? The 7-step recipe

Here is the full recipe, step by step. Each step maps to a specific recommendation in Anthropic's documentation, translated into plain language. You do not need all seven for a quick question. Use the first three for everyday work and all seven when you hand Claude something big.

Step 1: Open with the outcome, not the activity

Write the first sentence of your prompt as a description of the finished result. Not "help me with my pricing page" but "rewrite my pricing page so a first-time visitor understands what each plan includes within 30 seconds." The difference sounds small. It is not. A model that knows what finished looks like can check its own work against that picture. A model that only knows the activity cannot.

A useful test: could a stranger read your first sentence and tell you whether the job is done? If not, sharpen it before you send it.

Step 2: Give the reason, not only the request

Anthropic's guide is explicit that Fable 5 performs better when it understands the intent behind a request, and it offers a fill-in-the-blank pattern. Adapted to plain language:

I am working on [the larger task] for [who it is for].
They need [what the output enables].
With that in mind: [your request].

An illustrative example, not a client result: "I am preparing a budget review for my co-founder, who is skeptical about our software spend. She needs to see which subscriptions we actually use. With that in mind: turn this CSV export into a one-page summary ranked by cost per active user." The model now knows the audience is skeptical, the format is one page, and the ranking that matters is cost per active user. None of that was in the bare request.

Step 3: Set the boundaries out loud

Fable 5 is proactive. Anthropic notes it can occasionally take helpful but unrequested actions, like drafting an email nobody asked for. The fix is one plain sentence about what is in plan and what is not. Three boundary lines worth stealing:

• "If I am describing a problem or thinking out loud, give me your assessment. Do not make changes until I ask."
• "Do not add anything beyond what this task requires. The simplest version that works well is the right version."
• "Only pause to ask me when something is irreversible, changes the plan, or genuinely needs information only I have. Otherwise keep going."

That last line matters more than it looks. It tells the model exactly when interrupting you is correct, which means fewer needless check-ins and no silent guesses on the decisions that are actually yours.

Step 4: Pick the effort level deliberately

On the API, effort is the main dial for trading speed and cost against depth: low, medium, high, or xhigh. Anthropic recommends high as the default, xhigh for the most capability-sensitive work, and medium or low for routine tasks. The detail most people miss, straight from the documentation: lower effort on Fable 5 still performs well and often beats the maximum setting on prior models. You are not downgrading by choosing medium for everyday work. You are right-sizing.

Step 5: Demand evidence with progress reports

On long tasks, ask the model to tie every progress claim to something it actually did. Anthropic reports that in its testing, this single instruction nearly eliminated fabricated status reports, even on tasks designed to provoke them. A plain-language version you can paste:

Before you tell me something is done, check it against what you
actually did in this session. Only report work you can point to.
If something is not verified yet, say so plainly. If a step
failed or was skipped, tell me that too.

This is the difference between "all five sections are drafted" because the model checked, and the same sentence because it sounds plausible. On anything that runs longer than a few minutes, this instruction earns its keep.

Step 6: Ask for a readable summary, not a data dump

Fable 5 at high effort can explain more than you need. Anthropic's fix is a short brevity instruction rather than a list of formatting rules. The plain version:

Lead with the outcome. Your first sentence should answer
"what happened" or "what did you find." Keep the details that
change what I would do next, drop the rest, and write in
complete sentences rather than fragments and abbreviations.

Note what this does not say: it does not ask for "short" answers. Anthropic draws the distinction precisely. Readable beats brief. You want fewer topics, fully explained, not all topics compressed into shorthand.

Step 7: For big jobs, ask it to check its own work on a schedule

Anthropic recommends making self-verification explicit on long-running tasks, and notes that fresh-eyes review (a separate checking pass that did not write the work) tends to outperform the model grading its own homework in the same breath. The pattern to borrow: "Establish a way to check your work as you build, and run that check at regular intervals against the original goal." For multi-hour projects, also ask it to keep simple notes on what it learned, one lesson per note. Anthropic highlights that Fable 5 performs notably better when it can record lessons and reread them later.

Which effort level should you use for which task?

This table summarizes Anthropic's effort guidance as a quick decision aid. The use-case column is our plain-language interpretation for business work; the recommendation column follows the documentation.

Cost context for API users: at $10 per million input tokens and $50 per million output tokens, a long working session that reads 200,000 tokens and writes 20,000 costs about $3 in input and $1 in output. Treat those numbers as arithmetic on the published prices, not a quote; your usage will vary, and batch processing halves both rates.

When should Claude stop and ask you something?

The answer you encode in your prompt: only at real decision points. Anthropic's documentation describes the failure on both sides. Un-steered, a model can pause to ask permission it does not need, or push through a decision that was genuinely yours. The boundary instruction from Step 3 solves both, and the test below is the whole logic:

One related quirk Anthropic documents: deep into a very long session, the model can occasionally announce what it is about to do and then stop without doing it. If that happens, replying "go ahead and do it end to end" is enough. If you run automated pipelines, the documentation provides a standing instruction that prevents it entirely by telling the model to act on its own last paragraph before ending a turn.

Three copy-paste prompt templates

These templates apply the recipe. Fill in the brackets. They are written for the Claude app and work the same through the API. All bracketed scenarios are illustrative examples, not client work.

Template 1: the everyday task.

Outcome: [what finished looks like, one sentence].
Context: this is for [who], because [why it matters].
Boundaries: do not change [thing to leave alone].
If anything is ambiguous, make the sensible call and
flag it at the end rather than stopping to ask.

Template 2: the document or analysis job.

I am working on [larger task] for [audience].
They need [what the output enables].
With that in mind: [the request].
Lead your answer with the conclusion. Include only the
detail that would change my next decision. If the data
does not support a confident answer, say so directly.

Template 3: the long-running project.

Goal: [the outcome, with a definition of done].
Work through this end to end without waiting for me.
Pause only if something is irreversible, changes plan,
or needs information only I have.
Check your work against the goal at regular intervals.
Before reporting progress, verify each claim against what
you actually did. Unverified work is reported as unverified.
When finished, summarize: outcome first, then anything
you need from me, in complete sentences.

The 8-line cheat sheet: paste these, not prompt hacks

Most people prompt. Work orders get better results. This is the condensed version of everything above: the eight instructions we paste into our own working setups, copied from configurations we actually run, not theory. Each one is short enough to keep in a notes app and reuse on every serious task.

Here's a problem we haven't solved: [describe it]. plan it, ask clarifying questions, then execute end to end.

I'm working on [the larger task] for [who it's for]. They need [what the output enables]. With that in mind: [request].

When you have enough information to act, act. Give a recommendation, not an exhaustive survey.

Don't add features or abstractions beyond what the task requires. Do the simplest thing that works well.

Lead with the outcome. First sentence answers "what happened" or "what did you find." Supporting detail comes after.

If I'm describing a problem or thinking out loud, report your assessment and stop. Don't apply a fix until I ask.

Before reporting progress, audit each claim against a tool result from this session. Only report work you can point to evidence for. If it's not verified, say so.

Keep a notes file. One lesson per note, one-line summary on top. Record corrections and confirmed approaches, and why they mattered. Check it before new work.

Does this work in the Claude app, or only on the API?

Both. The recipe is the same wherever you type the prompt. The differences are mechanical. In the Claude app, you do not set an effort level with a parameter; the product manages depth for you, so Step 4 becomes a judgment about how much time you give the task rather than a setting you choose. On the API, effort, the 1 million token context window, and the 128,000 token output ceiling are explicit controls your developer (or your automation platform) configures per request.

If your team builds automations on the API, two of Anthropic's recommendations deserve a direct mention to whoever maintains them. First, requests on hard tasks run longer than they did on earlier models, so client timeouts and progress indicators need updating before you switch. Second, the documentation recommends giving long-running agents a dedicated way to send you messages mid-task, so updates reach you exactly as written instead of being summarized after the fact. Neither changes how you write prompts; both change whether the system around the model behaves the way you expect.

What are the most common Claude Fable 5 prompting mistakes?

Mistake 1: porting your old mega-prompt. Long rule lists written for earlier models can now hurt quality. Anthropic recommends reviewing old prompts and removing instructions that the model handles well by default. Try the short version first; add a rule back only when you see the specific failure it prevents.

Mistake 2: testing it only on easy tasks. Anthropic is unusually direct here: teams that test Fable 5 on simple workloads undersell it, and the best results come from assigning problems harder than what you would give earlier models. Pick something on your list that has felt too big, define the outcome, and hand it over.

Mistake 3: treating long turns as failure. A hard request at high effort can run for many minutes. Interrupting and re-sending teaches you nothing and wastes the work in progress. Decide the effort level up front, then let the run finish.

Mistake 4: asking it to show its private reasoning. Prompts that demand the model transcribe its internal thinking into the reply can trigger a refusal category on Fable 5, per Anthropic's documentation. Ask for conclusions, evidence, and the checks it ran instead. That gives you the trust signal without the refusal risk.

Mistake 5: vague success criteria. "Make it better" forces the model to guess your taste. "Make it readable for someone seeing it for the first time, and keep it under a page" is checkable. Every prompt should contain at least one sentence a stranger could use to judge whether the job is done.

Mistake 6: skipping the why. The request "summarize this contract" produces a generic summary. Adding "because I need to decide by Friday whether the cancellation terms are acceptable" produces the summary you actually needed. One clause of intent routinely outperforms a paragraph of formatting instructions.

The bottom line

Prompting Claude Fable 5 is closer to delegating to a capable colleague than to programming a machine. Say what done looks like, say why it matters and who it is for, set the fences, pick how hard it should think, and require evidence behind every progress claim. Anthropic's own documentation carries the headline insight: the model now follows brief, plain instructions well enough that clarity beats length every time. Write less. Mean more.

If you want help putting AI to work on the part of your funnel that is losing money, start with the free audit. Get a free audit.

Sources and method

This guide is based on Anthropic's first-party documentation: Prompting Claude Fable 5 (behavioral guidance, effort recommendations, instruction patterns, and the testing claims cited above), Introducing Claude Fable 5 and Claude Mythos 5 (availability, context window, output limits), and Anthropic's pricing page (per-token rates and the batch discount), all accessed June 10, 2026. Prompt snippets are paraphrased into plain language from the documented patterns. Worked examples in this post are labeled illustrative; they are teaching devices, not client results. Cost arithmetic applies the published per-token prices and is not a quote.

Related: AI Productivity vs revenue movement: Why Hours Saved Is Not a Business Metric · 10 ChatGPT Prompts That Transform Your Marketing Workflow · How to Map a Marketing Workflow in 60 Minutes