# Claude Fable 5.2 prompts: how to write them, with 5 ready-made examples

> How to write Claude Fable 5.2 prompts: effort, finish line, style rules and 5 examples for an HR office, based on Anthropic's official guide.

Source: Giallo Studio (giallostudio.it) · Page: https://giallostudio.it/en/blog/claude-fable-5-2-prompts · Published: 2026-11-09 · Updated: 2026-11-09

**An effective Claude Fable 5.2 prompt is shorter and more precise than those written for earlier models: it says what “done” means, names the actions that need a go-ahead, describes the style without hard limits, and leaves it to the effort setting to decide how much to reason.** The reason lies in the behavior of the model, released by Anthropic on 9 November 2026: it follows instructions literally, almost always thinks before answering and, on long tasks, may end its turn early.

This guide is based on [Anthropic's official guide to prompting Fable 5.2](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-2), consulted on 9 November 2026. The test results cited are reported by Anthropic and have not been replicated by third parties. For the model overview, see the article [Claude Fable 5.2: what it is and what changes](https://giallostudio.it/en/blog/claude-fable-5-2-what-it-is); for the basics of prompting Claude, see [Prompts for Claude](https://giallostudio.it/en/blog/claude-prompts).

<div class="key-points">
<p class="kp-title">In summary</p>
<ul>
<li>Fable 5.2 applies style instructions very literally: a limit such as “one instruction per sentence, in 20 words or fewer” can make the text choppy, according to the <a href="https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-2">Official Guide</a>.</li>
<li>A line asking it to “answer without reasoning” does not work at medium effort and above: for easy requests, lower the effort (<a href="https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2">“What's new” page</a>).</li>
<li>The default effort in the API is high, and Anthropic says most of the capability improvement is already present at that level.</li>
<li>On long tasks, especially at low and medium effort, the model may stop early: the fix is an explicit definition of “done” in the prompt.</li>
<li>Approval rules should name the exact actions: a rule that is too broad makes the model hold back routine work as well.</li>
<li>Change prompts in new conversations: editing the instructions mid-conversation can cause a 400 error in the API.</li>
</ul>
</div>

## What changes when you write a prompt for Fable 5.2?

Mostly, how much each word weighs. Fable 5.2 reacts to what the prompt says, and to what it leaves out, more than earlier models do. The [official What's new page](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2) describes the behaviors to know about, and the [prompting guide](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-2) explains how to respond. In a table:

| Observed behavior | What to do in the prompt |
|---|---|
| Thinks even when told not to (medium effort and above) | Remove the line and lower the effort for that type of request |
| Ends the turn before it has finished | Define “done” and ask it to get past ordinary obstacles |
| Applies style rules literally | Write a few rules, with the reason, and no hard numeric limits |
| Applies approval rules as written | Name the exact actions that need a go-ahead |
| Reacts to who “owns” the work | Say clearly whether it may change the structure or should touch as little as possible |
| Works silently in long sequences of tool calls | Ask for a one-line update now and then |

These are not flaws to work around with tricks. They are signs that the model does what you write, so the prompt has to be written in full and checked. The complete list of declared limits is in [Fable 5.2 known limitations and how to work around them](https://giallostudio.it/en/blog/claude-fable-5-2-limitations).

## Which effort level should you choose before even writing the prompt?

Start at high, which is the default in the API, and try dropping to medium where quality holds. The [Official Guide](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-2#calibrate-effort) distinguishes five levels by type of work:

- **low**: quick exchanges, well-specified tasks and high-volume work.
- **medium**: everyday knowledge work, chat, interactive coding.
- **high**: research, agentic tasks and long unsupervised runs.
- **xhigh**: only for the hardest work, when you have measured an advantage.
- **max**: for measuring the model's limit, not for normal use.

Effort changes the length of replies very little. To shorten or lengthen a text you need the prompt, not the effort. For the details of the levels, see the article [Fable 5.2: reasoning levels and effort](https://giallostudio.it/en/blog/fable-5-2-effort-levels); if you are coming from Opus, see [the reasoning levels of Opus 5.5](https://giallostudio.it/en/blog/claude-opus-5-5-reasoning-levels).

## How do you stop Fable 5.2 from stopping halfway through the work?

You write into the prompt what “done” means and what to do when facing an obstacle. The [Official Guide](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-2#finish-the-whole-task) provides a text to add to the system prompt, which asks the model not to stop while work is still owed, to diagnose errors and timeouts before giving up, and to leave deliberate blocks alone, such as a file marked “do not touch”. In Anthropic's test on recorded sessions at medium effort, with a version of that text the sessions that ended with doable work left roughly halved, but output tokens rose by about 30%.

Two practical consequences. First: the fix costs money. Second: because the model asks for confirmation less often, **keep a confirmation step of your own for risky or irreversible actions**, such as deleting data or sending messages to outside parties. The same guide says so explicitly.

<div class="callout"><strong>Warning</strong> The “do not stop” text does not replace human approval. If the task touches employee data, contracts or external communications, write in the prompt which actions need your go-ahead and which do not.</div>

## Why do style rules need to be written with care?

Because Fable 5.2 takes them literally. The [official page](https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2) warns that a hard limit on sentence length can make prose choppy. The same applies to approval rules: in an Anthropic test, a rule that listed “sending messages to people outside this conversation” among the steps needing approval made the model hold back roughly half of its routine requests to colleagues and other teams.

In practice, the difference shows right away:

- **Avoid:** “Write sentences of no more than 15 words. Never more than one idea per sentence.”
- **Better:** “The reader is an employee who has ten seconds. Short sentences where possible, a friendly and direct tone, no legal jargon.”

The second describes the result you want; the first imposes a form that the model will respect even when it makes the text worse. For working on writing in general, see [Claude AI to write better](https://giallostudio.it/en/blog/claude-for-writing).

## Five example prompts for an HR office

The examples that follow are **examples we built** from the official guidance, for the HR office of a company with 80 employees. They are not real cases or Anthropic texts. Adapt them and test them on your own documents. Before entering candidate or employee data, read the privacy section below.

### 1. Summarizing job applications, at medium effort

```
You have 25 attached applications for the administrative assistant role.
For each one write: relevant experience, gaps against the requirements
in the job ad, questions to ask at the interview.
The reader is the HR manager: she has read the job ad, not the CVs.
Done means: 25 profiles, none skipped, in alphabetical order.
Do not give scores and do not reject anyone: the decision is ours.
```

The prompt says what to produce, for whom and when it is finished. The last line sets a clear limit: the model prepares, the person decides.

### 2. Replying to employees about leave and time off

```
Answer the employee's request using only the attached
regulations. If the regulations do not cover the case, say so clearly
and name who in HR can decide.
Tone: friendly and direct, like an experienced colleague. Short sentences
where possible.
Do not promise anything on behalf of the company. Show me the draft:
do not send it.
```

Note the style rule: it describes the tone, with no numbers. “Show me the draft” names the action that needs a go-ahead, instead of a generic formula such as “ask for confirmation on important things”.

### 3. A long task with a finish line

```
The attached folder contains the 80 employee contracts.
Fill in a table with: name, hire date, end of the
probation period, end date of any fixed-term contracts.
Done means: 80 rows, every row checked against the document,
doubtful cells marked as "to be checked".
If a file does not open or is unreadable, try another way before
flagging it. If a file is marked "confidential - do not process",
leave it alone and list it at the end.
Do not change the contracts and do not send anything.
```

This applies the distinction made in the official guide: ordinary obstacles are overcome, deliberate blocks are respected and reported. With a task this long, at low or medium effort the risk of ending early is higher, so the finish-line line is essential.

### 4. A final report management can follow

```
Write the summary for management on the internal climate survey.
The reader has not seen the data and is coming from other work.
In the first three sentences: what you found, what you recommend, what you need
from us. Then the details.
Use simple words, full sentences, the same term for the same
thing. Do not use acronyms you made up during the analysis.
```

This is a shortened version, in Italian, of the idea in the “Make reports easy to follow” section of the [Official Guide](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-2), which asks for self-contained texts for a reader who has not seen the work.

### 5. Fable as an advisor: asking for a second opinion

```
Situation: I am preparing a new remote work policy.
Decision to make: whether remote work applies to everyone
or only to some roles.
What I have already done: I have read the national collective agreement and the current
policy (attached).
What I am asking you: the three main risks of each option and what
I should check with the employment consultant before deciding.
Do not write the text of the policy.
```

When a smaller model does the work and consults Fable as an advisor, the request gets useful context if it contains the situation, the decision, the work already done and a precise question: this is our own practical approach, not a rule in the documentation. How the advisor role works in the API, with its rules, is covered in [Fable 5.2 as Advisor and Orchestrator](https://giallostudio.it/en/blog/fable-5-2-advisor-orchestrator).

## Which lines of old prompts should you remove?

The official guide lists instructions written for earlier models that, on Fable 5.2, may cost more or serve no purpose. In prompts you inherit from Opus 5.5 or Fable 5.1, check these:

<ul class="checklist">
<li>Lines such as “answer directly without reasoning”: they do not stop the thinking at medium effort and above.</li>
<li>Invitations to delegate to subagents, if the model launches more than it needs: subagents can have a large effect on the cost of a session.</li>
<li>A token budget written into the prompt by your system: Anthropic saw lower costs, with equivalent performance, after removing a large budget.</li>
<li>The two sentences about “treat the previous answer as closed”: they reduce thinking on later questions, but the model flags its own past mistakes far less.</li>
<li>Notes, memories and absolute rules accumulated over time: a “never” written months ago can make the model hold back routine work.</li>
</ul>

One last technical precaution. The official guide recommends making these changes **only in new conversations**: with Fable 5.2, changing the system prompt, tools or earlier messages after the model has already produced reasoning blocks can cause a 400 error in the API. If you only use the Claude app, you are not affected. If you have an integration, you will find the details in the article on the [API migration to Fable 5.2](https://giallostudio.it/en/blog/fable-5-2-api-migration).

## What should you know about privacy before pasting employee data?

The examples above involve personal data. According to the [Covered Models support page](https://support.claude.com/en/articles/15425695-covered-models) , consulted on 9 November 2026, Fable 5.2 requires retention for at least 30 days, and zero data retention is not available by default. For résumés, contracts or employee health data you therefore need a privacy notice, a data processing agreement and data minimization: replace names with codes where you can, and talk to your privacy consultant or DPO. The overview is in [Fable 5.2 and privacy](https://giallostudio.it/en/blog/fable-5-2-privacy-data-retention) and in [GDPR and AI in the company](https://giallostudio.it/en/blog/ai-and-gdpr). The use of AI in staff selection also has its own rules: [AI Act: obligations for SMEs](https://giallostudio.it/en/blog/ai-act-sme-obligations).

## Which mistakes waste time and money with Fable 5.2 prompts?

1. **Writing “don't reason too much”.** It changes nothing at medium effort and above, and leaves you with a longer prompt and a false sense of security. Lower the effort instead.
2. **Setting hard numeric style limits.** Literal compliance produces jerky text. Describe the reader and the tone.
3. **A broad approval rule.** “Ask for confirmation before contacting outside people” stops routine emails too. List the actions, one by one.
4. **Using Fable 5.2 for everything.** It costs $8 and $40 per million tokens against $4 and $20 for Opus 5.5, according to the [official page](https://platform.claude.com/docs/en/models/fable-5-2/overview), and Anthropic recommends starting with Opus. For the comparison, see [Fable 5.2 vs Opus 5.5](https://giallostudio.it/en/blog/fable-5-2-vs-opus-5-5).
5. **Trusting only the final summary.** On coding tasks the official guide recommends reviewing the actual changes, not just the model's summary, which may leave out the small ones.

## How do you introduce these prompts in an office without improvising?

<ol class="steps">
<li><strong>Pick 10-15 real tasks</strong> from the office, with documents already anonymized.</li>
<li><strong>Write the prompt with the structure shown above</strong>: reader, finish line, actions that need a go-ahead, style described in words.</li>
<li><strong>Try medium and high effort</strong>, on Opus 5.5 and on Fable 5.2, comparing quality, time and cost of each run.</li>
<li><strong>Save the prompts that work</strong> in a shared document, with the date and the model, and reread it at every version change.</li>
</ol>

For prompts designed to work across several models, useful if you do not want to tie yourself to a single version, see [Prompts that work across multiple models](https://giallostudio.it/en/blog/prompts-that-work-across-models); for very long documents, see [Claude for long documents](https://giallostudio.it/en/blog/claude-for-long-documents); if you have used Opus, see [Prompts for Claude Opus 5.5](https://giallostudio.it/en/blog/claude-opus-5-5-prompts).

## Which prompt should you try first?

The best prompt for Fable 5.2 is one you can explain in a sentence: who is reading, what done means, what needs a go-ahead. If you want to set up a prompt library for HR or administration, tested on your own documents, and work out whether you need Fable or Opus is enough, [Giallo Studio's AI consultancy](https://giallostudio.it/en/ai-consulting) starts with a test on your real cases before proposing a model.

## Frequently asked questions

### How do you write a good prompt for Claude Fable 5.2?

By saying what “done” means, which actions need a go-ahead and what style you want, and nothing more. Fable 5.2 follows instructions literally, so every extra line has an effect. Anthropic's official guide recommends starting at high effort and testing every change on your own cases.

### Can I tell Fable 5.2 not to reason so it answers faster?

No, a single line like that is not enough. In Anthropic's tests, at medium effort and above the model kept thinking even with the instruction “answer directly”. For simple requests you lower the effort level; you do not write it in the prompt.

### Why does Fable 5.2 sometimes stop before it has finished?

It happens mostly on long, multi-step tasks and at low or medium effort: the turn ends with a list of what remains or an offer to continue. You reduce it by writing in the prompt not to stop while work is still owed and to get past ordinary obstacles, while still keeping confirmation for risky actions.

### Should style rules go in the prompt?

Yes, but only a few, with a clear purpose. Fable 5.2 applies them literally: a hard limit such as “one instruction per sentence, in 20 words or fewer” can make the text choppy. It is better to describe the reader and the tone, and keep numeric limits only where they are really needed.

### Are old prompts written for Opus 5.5 or Fable 5.1 still fine?

Reread them before reusing them. Some instructions written for earlier models, such as invitations to delegate to subagents or to think less, increase costs, have side effects or no longer do anything. Change them in new conversations, not in ones already under way.

### Do you really need Fable 5.2 to write office prompts?

Almost never. Anthropic recommends starting with Claude Opus 5.5 and moving to Fable 5.2 only when Opus, even at higher effort, is not enough. The prompts in this article work as a starting point on Opus too, but test them on the model you will actually use.
