10 copy-paste prompts built for Claude Fable 5 — long-document analysis, deep research, code review, ad scripts, and a meta-prompt that upgrades all the rest.
Claude Fable 5 rewards a different kind of prompt. The tricks that squeezed performance out of older models — "think step by step", ALL-CAPS commands, temperature tuning — are either built in now or gone entirely. What works on Fable 5 is precision: the full task up front, a concrete picture of what done looks like, and explicit permission to make small decisions without asking you.
We have been running Anthropic's new flagship through our daily workflows at PromptsRush since it landed — research briefs, code reviews, ad scripts, 500-page document dumps. The 10 prompts below are the ones that survived. Each one is copy-paste ready, with bracketed placeholders and a note on why it is shaped the way it is.
If you want the full background on the model itself, read our breakdown of Claude Fable 5 and Claude Mythos 5 first. This post is purely about getting the most out of it.
Four things changed with this model, and all four change how you should write prompts.
| Spec | Claude Fable 5 |
|---|---|
| Position | New top tier, above Claude Opus 4.8 |
| Context window | 1 million tokens (roughly 2,000+ pages) |
| Max output | 128K tokens |
| Thinking | Adaptive — the model decides how hard to reason per request |
| Temperature / top_p | Removed — steering happens in the prompt |
| API pricing | $10 / $50 per million input / output tokens |
It follows instructions literally. Fable 5 will not silently generalize one instruction to a similar case, and it will not infer requests you did not make. That sounds like a limitation. In practice it is the upgrade: prompts behave predictably, the same way every time. But it means vague prompts get vague results — the model is no longer papering over your underspecified asks.
There is no temperature dial. Anthropic removed sampling parameters on its newest models. If you want variety, creativity, or restraint, you ask for it in words. Every prompt below bakes that steering in.
It decides its own thinking depth. Adaptive thinking means "think step by step" is dead weight. Instead of telling the model how to reason, you tell it what a finished answer looks like and let it allocate effort.
The context window is enormous. One million tokens means entire codebases, complete contracts, or a year of customer feedback in a single message. The biggest prompting mistake we see is drip-feeding context across ten turns when one big paste would do.
Pro tip: Fable 5 is more deliberate than older Claude models — on minor choices it tends to pause and ask. Adding one autonomy line ("pick a reasonable option and note it instead of asking") cuts the back-and-forth dramatically. You will see that line reused across these prompts. It is doing real work.
This is the master template. Fable 5 performs best when the entire job arrives in one well-specified message rather than dripped out over a conversation — long-horizon work is exactly where this model pulls ahead of everything else.
You are acting as my [ROLE — e.g. senior data analyst]. TASK: [Describe the complete task in 2-4 sentences. Include everything up front.] CONTEXT: [Paste relevant background, data, or files here.] CONSTRAINTS: - [Constraint 1 — e.g. keep it under 800 words] - [Constraint 2 — e.g. use only the data I provided] DONE LOOKS LIKE: [Describe the finished output concretely — format, length, sections, file type.] For minor decisions (naming, formatting, which of two equivalent approaches), pick a reasonable option and note it instead of asking me. Only stop to ask if the scope itself is unclear.
The DONE LOOKS LIKE block is the part most people skip and the part that matters most. "Write a report" produces a generic report. "A 600-word memo with a 3-bullet summary on top, written for a CFO" produces the thing you actually wanted.
The 1M context window is Fable 5's most underused feature. Most people summarize documents in chunks because that is what older models forced. Stop chunking — paste the whole thing and ask for cross-document contradictions, which chunked workflows can never find.
I am pasting [DOCUMENT TYPE — e.g. a 200-page contract / our complete codebase / 12 months of customer feedback] below. Read all of it before answering. Do not skim or answer from the first section alone. Then give me: 1. A 5-bullet executive summary of the whole document 2. The 3 sections that most affect [MY GOAL — e.g. our renewal decision] 3. Anything that contradicts itself across different sections — quote both passages 4. The single most important thing I would miss on a quick read [PASTE DOCUMENT HERE]
Point 3 is the killer feature. We ran a vendor contract through this and it surfaced a termination clause in section 14 that quietly contradicted the renewal terms in section 3. No chunked summary catches that.
Fable 5 is a noticeably better thought partner than its predecessors — more willing to commit to a position and to tell you its confidence level honestly. This prompt forces that behavior into a structure you can act on.
Research question: [YOUR QUESTION] Build me a decision-ready brief: 1. Current state of the topic — what is established vs. still debated 2. The 3 strongest arguments on each side, steelmanned 3. What the evidence actually supports, with your confidence level (high / medium / low) per claim 4. What would change your conclusion if it turned out to be true Rules: separate facts from your inference and label which is which. If you are not sure about something, say so directly instead of hedging with vague language. End with a one-paragraph recommendation written for a busy executive.
For research that needs live web sources and multi-step agent workflows on top of the model, we pair prompts like this with Genspark — our full take is in the Genspark review.
Fable 5 finds more real bugs than any model we have tested — but it also follows severity filters literally. Tell it "only report important issues" and it will silently drop findings it judged below the bar. So we tell it the opposite: report everything, rank later.
Review the code below as a senior engineer doing a pre-merge review. Report EVERY issue you find, including ones you are uncertain about or consider low severity. Do not filter for importance — surfacing a finding that gets dismissed is better than silently dropping a real bug. For each finding give: - File/line or function name - What is wrong and the failure it could cause - Severity (critical / major / minor) and your confidence (high / medium / low) - A suggested fix in code Finish with: the 3 findings you would fix first, and why. [PASTE CODE OR DIFF HERE]
If you write code with Claude daily, our 100 Claude Opus 4.7 prompts for power users and Next.js prompt collection both transfer to Fable 5 nearly unchanged.
Most people use AI to validate decisions they have already made. This prompt does the opposite — and Fable 5 is the first model we trust with it, because it actually pushes back instead of folding into agreement after one objection.
I am about to make this decision: [DESCRIBE DECISION AND YOUR CURRENT PLAN] Do not validate my plan. Your job is to stress-test it: 1. The 3 strongest reasons this fails — be specific, not generic 2. Which of my assumptions is doing the most load-bearing work, and what happens if it is wrong 3. The argument a smart skeptic would make against this in one paragraph 4. What I should verify in the next 7 days before committing Then, and only then, tell me whether you would proceed — and push back on me if you would not.
Older models needed prefilled responses or fragile regex cleanup to return clean JSON. Fable 5 follows an explicit schema description literally — which makes this prompt close to deterministic.
Extract structured data from the text below. Return ONLY a valid JSON array — no markdown fences, no commentary. Each object must have exactly these fields: - "name" (string) - "category" (one of: [LIST YOUR CATEGORIES]) - "value" (number — use null if not stated, never guess) - "source_quote" (string — the exact sentence you extracted it from) If a field is not present in the text, use null. Do not invent values. [PASTE TEXT HERE]
The source_quote field is your audit trail. When a value looks wrong, you check the quote instead of re-reading the source document. It also measurably reduces invented values — the model will not fabricate a number it has to attach a real sentence to.
Fable 5 writes warmer and cleaner prose than any previous Claude, with far fewer AI tells. That makes it the first model worth trusting with your own voice — if you show it your voice first.
Below are 3 samples of my writing, then a draft that needs rewriting. First, describe my voice in 5 specific attributes (sentence length, formality, vocabulary, rhythm, how I open and close). Then rewrite the draft to match my voice exactly. Keep every fact and claim intact — change only the delivery. Do not smooth out my opinions or add hedging I would not use. MY SAMPLES: [PASTE 2-3 WRITING SAMPLES] DRAFT TO REWRITE: [PASTE DRAFT]
The describe-first step matters. Forcing the model to articulate your voice before rewriting produces a noticeably closer match than "rewrite this in my style" — and the 5-attribute description it generates is reusable in future prompts.
Ad copy is where literal instruction-following pays off twice: Fable 5 actually respects "no marketing words" instead of sneaking in a "game-changer" by paragraph three. Feed the winning script into Arcads to turn it into a finished UGC video with AI actors.
Write 5 UGC-style video ad scripts for [PRODUCT] targeting [AUDIENCE]. Each script: 30-45 seconds, structured as HOOK (first 3 seconds, spoken to camera), PROBLEM (relatable and specific), PRODUCT MOMENT (one concrete benefit shown, not listed), CTA (casual, not salesy). Make the 5 hooks genuinely different angles: a confession, a contrarian take, a before-and-after, a mistake I made, and a question. Write like a real person talks — contractions allowed, no marketing words like "revolutionary" or "game-changer".
Run 5 hooks, kill the bottom 3, double down on the winner. For the full workflow, see our guide to the best AI prompts for ads and commercials.
The fastest way we know to learn a new domain. The three-pass structure stops the model from defaulting to one middle-depth explanation that serves nobody.
Explain [TOPIC] to me in three passes: PASS 1 — One paragraph a smart 12-year-old would understand. No jargon. PASS 2 — One page for a professional in an adjacent field. Introduce the necessary technical terms and define each one the first time it appears. PASS 3 — The expert version: the 3 things practitioners actually argue about, where the field is heading, and the one misconception even informed people hold. End with a 5-question quiz I can use to test whether I understood it.
The meta-prompt. Paste any prompt you use regularly and have Fable 5 rebuild it for its own strengths. This is the fastest way to migrate a prompt library from older models.
Here is a prompt I use regularly: [PASTE YOUR PROMPT] Rewrite it specifically for Claude Fable 5: 1. Make every instruction explicit — Fable 5 follows prompts literally and will not guess intent 2. Add a concrete "done looks like" description of the output 3. Move any buried requirements up front 4. Add an autonomy line so it picks reasonable defaults instead of asking about minor choices 5. Cut any instructions that contradict each other Return the rewritten prompt in a code block, then list what you changed and why in 5 bullets or fewer.
The prompts compound. Our standard content pipeline runs four of them back to back, in one conversation so the context carries through:
Total time: about 20 minutes for work that used to take an afternoon. The same chain works for strategy memos, landing pages, and video scripts — swap step 4 for the UGC Ad Script Engine (#8) when the output is an ad.
Every prompt above is built from the same seven patterns. Steal them for your own prompts:
These templates are the starting set — the Prompt Upgrader (#10) will grow them into a library tuned to your work. For more on the model and the wider Claude ecosystem:
Browse our full prompt library, check the latest AI model profiles, or head back to the blog for more workflows like this one.
8 questions answered
Claude Fable 5 vs Opus 4.8 vs GPT-5.5: Benchmarks (2026)
Jun 9 · 7 min
Claude Fable 5 and Claude Mythos 5: Everything You Need to Know
Jun 9 · 10 min
Best Midjourney Prompts for Creators in 2026
Jun 8 · 8 min
How to Write Better design.md Files
Jun 8 · 7 min
AI Skills vs AI Agents: What's the Difference?
Jun 8 · 7 min