7 Prompting Lessons From the Leaked Fable 5.1 System Prompt

When a new frontier model ships, a copy of its system prompt tends to turn up within a day or two. Fable 5.1 was no exception.

A system prompt is the standing instruction set a company writes for its own model — the thing the model reads before it reads you. It gets tested harder than any prompt you will ever write: more cases, more users, more money riding on the result. That makes it worth reading as craft.

The striking thing about this one is where the length goes. Very little of it is phrasing or formatting. Most of it sets decision policy — what standard of judgment to apply, what the model may resolve on its own, which source wins when two of them disagree.

That is a lever almost nobody pulls in an everyday prompt. Seven of those decisions transfer directly, as single sentences you can bolt onto what you already write.

First, what this actually is

Worth being straight about the source before going further.

• Anthropic has not published this file. It is a copy that circulated after the release.
• It has not been verified as genuine — not here, and not anywhere else worth pointing you to.
• What follows is a reading of it, not a quotation. The file itself is not reproduced here.

So treat every “the prompt reportedly does X” below as exactly that — reported, second-hand, unconfirmed.

Here is why the post still holds up. None of the seven moves depend on the leak being real. Each one is a sentence you can add to your next prompt and judge for yourself in about thirty seconds. And four of the seven match prompting advice that was already well established from completely unrelated sources — creators, published vendor documentation, and in a couple of cases articles already on this site.

The leak is where these came from. Their value does not rest on it.

1. Name the kind of judgment you want

Most prompts specify a topic. Very few specify a job.

“What do you think of this?” is a topic. The model has to guess whether you want an explanation, a critique, a recommendation, or reassurance. The safe guess is a balanced overview with something encouraging on the end, so that is usually what you get.

The reported prompt is unusually specific on this point for its own model: answer the question first, be willing to push back, prioritise what is relevant over what is complete. Those are standards of judgment. None of them are topics.

You can set the same kind of standard in one line:

Evaluate this rather than explaining it. Give me your actual recommendation, including anything you think I’m overlooking.

I’m looking for a sanity check, not validation. If my premise is wrong, tell me.

Tell me what you’d do in my position — and what would have to be true for that to be the wrong call.

When it earns its place: any question where a balanced overview would annoy you. Purchases, career moves, whether a plan survives contact with reality, anything where you have already done the thinking and want a second opinion rather than a briefing.

This is also the cheapest defence against the thing models do worst, which is agree with you. There is a heavier version of that fix — handing the work to a second model to review — and it is worth the effort when the stakes justify it. One sentence is worth having as a reflex in between.

A model will give you a call if you ask for one. Ask for a summary and a summary is what arrives.

2. Give it permission to assume and keep going

Ambiguity leaves a model two options, and both are bad.

It can guess silently, and you find out later that the whole answer rests on something you would have corrected in four words. Or it can stop and ask, which is fine once and maddening by the third round trip on a task you thought you had handed over.

There is a third option, and it takes one sentence to unlock:

Don’t stop for minor clarifications. Make reasonable assumptions, state any important ones, and proceed.

Fable is reportedly instructed to work an ambiguous request as far as it can before asking for clarification, and to state its assumptions and carry on when the user has already supplied real constraints. You can ask for that behaviour directly.

When it earns its place: long prompts, and especially spoken ones. If you dictate — and dictation is one of the better habits you can build with these tools — you produce a lot of useful context and a handful of small ambiguities you genuinely do not care about. Without this line, those ambiguities are where the model stops.

What you get back is the work, plus a short list of the assumptions it rested on. Scanning that list takes ten seconds and catches the one assumption that actually mattered.

The assumptions were always there. This sentence gets them written down.

3. Say which sources win

The moment a prompt carries attached material and the model can search the web, you have a conflict you have not resolved: when your document disagrees with the internet, which one wins?

Most people never say. They paste everything in and hope the weighting works itself out.

The reported system-prompt version handles this hierarchically, by the kind of question being asked — internal and personal sources for questions about the user’s own situation, external sources for external facts, both together for comparisons. That structure is worth stealing even if the file turns out to be fabricated, because the underlying idea is right: precedence belongs to a type of claim, not to a document.

In practice that looks like:

Treat the transcript and the documents I’ve given you as authoritative about our product. Use the web only for outside context.

Use my figures for my situation. Use current external data only for things like tax rules and market conditions.

Base your answer primarily on this paper. If you add general knowledge, mark clearly which parts came from where.

When it earns its place: anything where you brought your own material and the model can also search. Research, financial questions, work documents, competitive analysis.

That last phrasing does something extra worth noticing. Asking the model to separate what your source said from what it added turns a single blended answer into two things you can check independently. The parts drawn from your document you can verify quickly. The parts it supplied from general knowledge are where the errors live, and now they are labelled.

4. Say what shape the answer should be — when you can picture it

The reported prompt names specifying length and format as an effective technique, which is the least surprising item on this list. It is worth including anyway, because the useful version is smaller than most people assume.

You do not need elaborate formatting instructions. These are enough:

Give me the three to five things that actually matter.

Start with your recommendation, then explain it.

Assume I’m intelligent and completely new to this.

Compare these in a table, then tell me which you’d choose.

When it earns its place: when you can already picture the output you want. That is the whole test. If you cannot picture it, specifying the shape guesses on your behalf and you lose the chance to be surprised by a better structure.

On casual questions, skip it. Current models format well enough unaided, and the instruction costs more attention than it returns.

If you want the fuller treatment of how instructions, context and constraints fit together, the basics of writing AI prompts covers the structure this sits inside.

5. Show it one good example and one bad one

For anything subjective — tone, style, voice, taste — examples outperform description by a wide margin. The reported prompt leans on them constantly to pin down behaviour that would be ambiguous if it were only described.

The move most people miss is the second example. Not the one showing what you want. The one showing what you do not.

Rather than a hundred words describing the tone you are after:

I want dry and funny, along the lines of: “Whoever decided porous grout belongs in a shower has some explaining to do.” Not corny, not an elaborate joke, and don’t make it sound AI-written.

Or for work:

Write this the way I’d actually say it in a meeting — concise, knowledgeable, slightly conversational. Don’t turn it into marketing copy.

When it earns its place: rewriting, drafting in your own voice, anything where you would recognise the right answer instantly but struggle to specify it in advance.

The anti-example does more work than the positive one. “Don’t make it sound AI-written” rules out a large region of output space that no amount of positive description reliably avoids.

This is also the documented fix for vague adjectives generally. Words like detailed, thorough and professional carry a definition in your head that the model does not share — “detailed” might mean one page to you and thirty-five to it. Define the word or show an example. Description alone leaves the gap open.

6. Name the date when the facts can go stale

Every model answers from a training snapshot unless something makes it go and look. The reported prompt is emphatic on this: named products, model versions, services, prices, anything current gets verified rather than recalled. The reasoning behind it is sound — a confident stale answer does more damage than an obvious admission of uncertainty, because you cannot tell it is wrong by looking at it.

Adding the word “latest” somewhere in your prompt does not do this. Naming a date does:

Use current sources where this could have changed. Don’t rely on training knowledge for current facts.

Research the current state as of September 2026 before making a recommendation.

When it earns its place: pricing, model comparisons, product recommendations, laws and rules, anything with a version number, anything competitive. AI tooling especially — this field invalidates its own documentation every few weeks.

Most chat tools now browse by default for questions that obviously need it. The failure case is the question that does not obviously need it: the one where the model does not realise its snapshot has expired, so it never thinks to check.

7. Ask for the distilled conclusion

This is the one place worth disagreeing with the reported file, which apparently repeats the standard advice that encouraging step-by-step reasoning improves results.

That advice has aged. Reasoning models already reason internally before answering, and how hard they think is now a setting — Claude’s effort level, ChatGPT’s thinking level — rather than something you unlock with a phrase. “Think step by step” mostly buys you a verbose reconstruction of reasoning you were never going to read.

What you actually want is the thinking without the transcript:

Think this through carefully, then give me the distilled conclusion and the key reasons.

Consider the competing explanations before answering. Give me your conclusion, your confidence in it, and the two to four factors driving it.

When it earns its place: hard decisions with more than one plausible answer. The second version is the stronger one, because asking for competing explanations forces the model to generate alternatives before committing, which is the part that improves the answer.

Here is the detail that makes this worth trusting more than the rest of the list. This same conclusion — that effort words have stopped earning their place, and that the compute lever is a setting now — was reached independently by working practitioners with no access to any leaked file, on the basis of watching output quality change across model releases. Two sources that never consulted each other landing in the same place is stronger evidence than either one alone.

That convergence is worth more than the leak is.

A default you can paste when the answer matters

You do not need a mega-prompt. Pasting a wall of standing instructions into every conversation makes your prompting worse, not better — it buries the specifics of the actual question under boilerplate the model has to work past.

But for the handful of questions each week where you genuinely care about being right, a short suffix does most of the work of this article:

Give me your actual assessment, not just an explanation or validation.
If my premise seems wrong, push back.

Use current information where it matters. Make reasonable assumptions rather
than stopping for minor clarifications, and call out any assumption that
materially affects the answer.

Lead with the conclusion, then the few reasons that matter most.

And when you have supplied your own material:

Treat the material I've provided as the primary source for what it says.
Don't silently fill gaps with general knowledge. If outside or current
information would materially improve the answer, go and get it — and keep
it clearly separate from what my source actually says.

Neither of these is clever. They are the moves above, written down once so you stop having to remember them.

Quick reference

Where these fall over

The assumption licence needs a real task behind it. Give a model permission to proceed on a prompt that was genuinely underspecified and it will proceed confidently in a direction you did not want. The line works because it is attached to context that constrains it. On a two-sentence request it just removes the safety check.

Asking for a judgment does not make the judgment good. “Tell me what you’d do” produces a recommendation every time. Whether it is a sound recommendation depends entirely on whether the model has the information to make one, and it will rarely volunteer that it does not. The sentence changes the shape of the answer, not the quality of the underlying reasoning.

Source precedence only works if you actually supplied the source. Telling a model to treat your document as authoritative when you have pasted two paragraphs of it gives you a confident answer built on two paragraphs.

All seven of these age. These are behaviours of current models, and the entire point of move seven is that a prompting rule everyone repeated last year is now dead weight. Whatever standing instructions you write down, expect to prune them on a schedule — the same maintenance problem that hits instruction files applies to your own prompting habits.

And the source is what it is. An unverified leak, read second-hand. If it turns out to be somebody’s invention, the seven moves still work, because none of them were ever taking the file’s word for anything.

Key takeaways

• Most of a frontier system prompt’s length goes on decision policy — how to judge, what to assume, which source wins — rather than on phrasing.
• The highest-leverage thing you can add to a prompt is one sentence naming the job you want done with the context you just supplied.
• Ask for a call rather than a summary, and you will get one.
• Give the model permission to assume and proceed, and ask it to state the assumptions. The assumptions existed either way.
• When you bring your own material, say which claims it wins on.
• “Think step by step” has been replaced by an effort setting. Ask for the conclusion instead.
• You do not need a longer prompt. You need a better final sentence.

Related reading:

• Master the Basics: How to Write AI Prompts — the structure these moves attach to
• Context Engineering: How to Get Smarter AI Outputs Every Time — the context-dumping half this post assumes you already do
• Don’t Let an AI Grade Its Own Homework — the heavier version of move one
• The Prompt That Keeps Your CLAUDE.md Tight and Current — pruning instructions that have stopped earning their place
• 25 Game-Changing ChatGPT Prompts According to Reddit Power Users — a starting library if you want prompts rather than principles