Claude Fable 5 Prompts for Developers & YouTubers: 12 Copy-Paste Wins
Twelve battle-tested prompts for Anthropic's narrative-tuned Claude Fable 5 — six for developers (PRs, commits, docs, refactors, error messages) and six for YouTubers (hooks, titles, scripts, descriptions, comment replies, channel art).
AI Magic
Editorial Team
June 11, 202610 min read46views
Anthropic's Claude Fable 5 is the narrative and language-craft model in the Claude lineup — tuned for tone, voice consistency, and the specific kind of writing where the words have to actually land. Opus 4.8 reasons. Haiku is fast. Fable 5 writes.
That makes it a perfect tool for two audiences most prompt libraries ignore: developers (who spend half their day writing — PR descriptions, commit messages, docs, refactor proposals, error messages) and YouTubers (whose entire output is written first, even when it ends up spoken). What follows is twelve ready-to-paste prompts — six per audience — that we use in production and refine constantly. Replace the {{placeholders}}, send to Fable, ship.
Fable 5 at a glance
Where it sits in the Claude lineup, May 2026
Fable 5
Anthropic's narrative model
Tuned for voice, tone, and prose that lands
~3x
Faster than Opus
On standard generation, no thinking mode
~$1.50
Output per 1M tokens
Roughly 1/10th of Opus 4.8 output pricing
500K
Context window
Same as Opus — Projects load fully
Why Fable 5 Specifically
Most Claude users default to Opus 4.8 for everything. For pure prose work, Fable beats it on three dimensions: it matches voice references more tightly, it produces less of the "AI tells" that pattern-detectors and readers catch (over-balanced sentence rhythm, the em-dash habit, the "not just X but Y" construction), and it costs roughly a tenth of Opus output. For the dozens-of-prompts-a-day rhythm that both developers and YouTubers actually run, that price gap matters.
Fable still loses to Opus on hard reasoning, math, and any task where the answer has to be correct more than it has to be well-written. The right pattern is to route: Fable for prose, Opus for the parts that need to think. Set up a router if you're hitting the API at scale.
How to Use These Prompts
Replace every {{placeholder}}. Don't leave them blank "to see what happens" — Fable handles ambiguity badly compared to Opus.
Drop sample voice into a Project. Your last 3 PR descriptions / last 5 video scripts / brand-voice doc. Fable picks up tone from context aggressively.
Don't add "you are a helpful assistant" preambles. Fable is already direct. Prefaces just bloat the cost.
Push back when the first answer isn't quite right. Fable updates voice cleanly when you say "this reads too formal" or "drop the second paragraph and rebuild."
Six Prompts for Developers
Six daily-workflow prompts that turn the prose layer of dev work — names, commits, PRs, docs, error messages — into something you'd be happy to put your name on.
Prompt
DEV-1 · Naming generator
You are helping me name a {{thing: function | variable | module | class | flag | service}}. Here is the context:
{{paste 5–15 lines of surrounding code, or 2 sentences on what it does}}
Generate 10 candidate names. For each name, give:
- The name itself
- A 1-line reason it fits
- Whether it still works if the scope grows (yes/no)
Constraints: match the {{language: JS | Python | Rust | Go | Swift | etc.}} naming conventions. Prefer concrete nouns and verbs over jargon. No more than 3 words per name. Avoid the suffixes 'Manager', 'Helper', 'Util', 'Handler' unless they are genuinely the most accurate label.
End with your single top pick and one sentence on why it wins.
Read this diff and write a PR description for it.
Diff:
{{paste your git diff or describe the changes}}
Context: {{1–2 sentences on why this work is happening — link the underlying issue, customer ask, or engineering decision}}
Output format:
## Summary
A 2-sentence summary: what changed, and why.
## What changed
- Bulleted list of the meaningful changes (not every line)
## Why
The motivation. What problem this solves. Reference the underlying issue or decision.
## Test plan
What to verify before merge. Cover the happy path and 1–2 edge cases worth checking.
## Notes for reviewers
Anything that needs human judgment. Tradeoffs taken. What you would tackle next.
Tone: direct, no fluff. Assume the reviewer is technical and busy.
Write a commit message for this change.
Diff:
{{paste diff}}
Format:
- Subject line: imperative mood, under 70 characters, lowercase except proper nouns
- Blank line
- Body: 2–4 short paragraphs explaining why the change is being made, what changed at a high level, and any risk worth flagging
No emojis. No 'added new feature'-style summaries. The reader is reviewing 50 commits today; help them parse this one in 10 seconds. If the diff is purely cosmetic (rename, formatting), say so in the subject line and keep the body to one sentence.
Convert this code into reader-friendly documentation.
Code:
{{paste function, class, or module}}
Audience: {{junior dev joining the team | senior dev who needs a fast overview | external API user}}
Output:
- A 1-sentence summary that fits in an autocomplete tooltip
- Parameters: each with type, default, and what it does semantically (not just what its type is)
- Return value: shape and what it represents
- 1–2 worked examples — minimal but realistic, not foo(1, 2)
- A 'Common pitfalls' section, but only if the function has genuinely non-obvious behavior
Tone: precise, no marketing. If a parameter is dangerous, say so plainly. If the function has a quieter alternative, mention it once.
This code works but it is bothering me:
{{paste 20–60 lines of code}}
What is bugging me: {{1 sentence — 'too long', 'hard to test', 'duplicates X', etc.}}
Produce a structured refactor proposal:
## Problem
Name the specific issue in plain English. Not a vibe — a concrete failure mode.
## Proposed refactor
Show the rewritten code or describe the structural change. Keep the diff scope minimal.
## Tradeoffs
What we lose. What gets harder. Honest list, never zero.
## Risk
What could break. What tests would need to exist.
## Recommendation
Your honest call: do this now, defer, or leave it alone.
Default to 'leave it alone' if the refactor is marginal. You are not being graded on always recommending change.
Rewrite this cryptic error message into something a user can act on.
Original error:
{{paste the raw error string or stack trace excerpt}}
Context: this surfaces in {{a CLI | a web app form | a logs panel | a developer dashboard}}.
Output 3 alternatives:
1. Strict — formal, suitable for a regulated product
2. Friendly — warm but not cute, suitable for a consumer app
3. Dev-focused — technical but with the immediate next step called out
Each alternative should:
- Name what went wrong in one short sentence
- Tell the reader exactly what to do next (one specific action)
- Avoid blame ('you forgot to…') and jargon the user cannot fix
End with your recommendation on which to ship, and one sentence on why.
Six prompts that cover the writing surfaces of a real channel — hooks, titles, scripts, descriptions, comment replies, and channel art copy. Pair them with ElevenLabs for voiceover, Hedra for talking-head B-roll, and the four affiliate CTAs surfaced on each prompt block below.
Prompt
YT-1 · Hook generator (8 variants)
Generate 8 hooks (0–7 second opener) for a YouTube video on this topic:
Topic: {{your video topic}}
Audience: {{who watches your channel — be specific: 'junior frontend devs', 'stay-at-home parents', 'indie game devs'}}
Channel voice: {{descriptive — 'blunt and tactical', 'warm and curious', 'fast and punchy'}}
For each hook, give:
- The line itself (under 20 words, spoken-word friendly)
- The hook archetype used (curiosity gap / contrarian take / stake reveal / numbered promise / face-to-camera question / counterintuitive fact / personal reveal / cliffhanger)
- A 1-line note on why it might land for this audience
Do not repeat archetypes across the 8. Each hook must work audio-only — assume the viewer has not yet seen the screen.
Generate 10 YouTube titles for this video.
Video: {{1 paragraph on what the video covers}}
Channel: {{channel name + 1 line of positioning}}
Recent best-performing title format: {{paste one of your last 5 winning titles for tone reference}}
Each title:
- Under 70 characters
- No emoji unless the channel typically uses them
- Front-load the most search-relevant noun phrase
- Avoid 'Ultimate Guide to…' and 'You Won't Believe…' — both trigger AI-generated pattern detection
For each title also output:
- The implied promise the viewer is clicking for
- Estimated CTR direction vs your recent baseline (higher / similar / lower) with a 1-line reason
End with your top 3 ranked, and which thumbnail concept each pairs with (face / object hero / before-after / number-on-screen / split-screen).
Build an outline for a {{8-minute / 12-minute / 20-minute}} YouTube video on:
Topic: {{topic}}
Audience: {{specific viewer profile, not 'everyone'}}
Goal: {{what should the viewer learn / feel / do by the end}}
Output:
- Opening (0:00–0:30): the hook, the stakes, what they will get
- Section 1 (0:30 to ~2:00): name it, key beats (3 bullets), B-roll cues
- Section 2 (~2:00 to ~5:00): name it, key beats, B-roll cues
- Section 3 (~5:00 to ~7:00): name it, key beats, B-roll cues
- Close (~7:00 to end): the payoff, the CTA, the next-video tease
Each section should:
- Have one clear point, not three
- Include 1–2 specific examples or numbers
- Note where a transition, pattern interrupt, or cutaway would help retention
Tone: {{your channel voice}}. Assume the viewer can skip — every beat earns its time.
Write a YouTube description for this video.
Video topic: {{topic}}
Key things covered: {{3–5 bullets}}
Target search terms: {{2–3 queries you want to rank for}}
Affiliate links or sponsors to include: {{paste links, or none}}
Channel positioning: {{1 line}}
Output:
First 2 lines (the part shown above 'Show more'):
- One sentence that promises the value
- One sentence with the primary search term naturally placed
Body (3–5 paragraphs):
- Expand on what the video covers, weaving in the target search terms naturally
- Do not keyword-stuff
- If there is a sponsor or affiliate disclosure, place it cleanly at the end of a paragraph, not as a list
Chapters:
- 4–7 realistic chapters with timestamps
- Each chapter title is searchable on its own ('How to [verb]…', 'The N mistakes…', etc.)
Close with 3–5 hashtags matching the search terms and a standard subscribe + next-video CTA.
Produce a comment-reply playbook for my channel.
Channel voice: {{1 line — 'blunt', 'warm', 'professorial', etc.}}
Niche: {{topic area}}
Common comment types I get:
- Generic praise ('great video')
- Specific praise ('loved the part about X')
- Sincere question
- Disagreement
- Off-topic / spam-adjacent
- Genuine criticism
For each type, give:
- A 1-line diagnosis of what is actually being asked or felt
- Two reply templates of different length (one sentence, and 2–3 sentences)
- A 'don't do this' note — the failure mode to avoid
End with one single rule for the channel's reply voice that I should keep consistent across every response.
Write copy for my YouTube channel banner.
Channel name: {{name}}
Niche: {{topic}}
Who it is for: {{specific viewer — not 'everyone interested in X'}}
What I uniquely promise: {{1 sentence}}
Posting cadence: {{e.g. 'weekly Sundays'}}
Produce:
- Primary headline (under 5 words, big-text-on-banner format)
- Subline (under 12 words, supporting promise)
- CTA line (under 6 words — 'New videos every Sunday' / 'Free templates in the description' / similar)
- Three alternate headlines with different angles: utility, transformation, contrarian
Then: a 1-line description for the 'About' tab that uses the search terms but reads like a human wrote it.
Tone: confident, specific. Banners that work commit to one promise; vague banners convert nothing.
Three patterns that turn these prompts from "one-shot helper" into something closer to a repeatable workflow.
Pair Fable with a brand-voice Project
Open Projects in claude.ai. Create one called "Voice — [your name / channel]." Drop in: your last 3 PR descriptions, your last 5 video scripts, a brand-voice doc, two examples of writing you've explicitly approved. Every prompt above runs inside this Project. Fable will pick up rhythm, vocabulary, and structural choices without you spelling them out.
Chain prompts — output of one is the input of the next
Run YT-3 (Script Outline), take its output, feed it into a fresh prompt: "Expand Section 2 into a 90-second voiceover script. Keep the same beats but write it as I would speak it." Take that script and pipe it into ElevenLabs for the voiceover, then into Hedra for the talking-head video. The whole loop is under 15 minutes.
Route hard reasoning to Opus, prose to Fable
If your prompt needs a correct answer (does this refactor actually work? does this regex match what I claim?), route it to Opus 4.8. If your prompt needs a well-written answer (turn this into a clean PR description, write this hook), route to Fable. Tools like LiteLLM and OpenRouter make the routing a single config change.
Feed Fable real source material, not just topics
Fable's output quality scales hard with input quality. For research-heavy YouTube scripts and dev PR descriptions where the facts matter, pipe in current source material via Firecrawl instead of relying on the model's training cutoff. One API call returns clean LLM-ready markdown from any URL — perfect Project fuel. Free tier · paid from $16/mo.
The right way to use Fable is as a writing partner with a fast cursor and a clean voice — not as an oracle. Bring the substance. Let it shape the words.
What to Do Next
Pick three prompts to try this week. The hook generator, the PR description writer, and the comment-reply playbook are the highest-leverage starting points for most teams.
Build a Voice Project. Spend 20 minutes assembling the brand voice context once. Pays itself back inside the first day.
Chain at least one workflow. Pick a single use case (e.g. video script → voiceover → talking-head clip) and run it end-to-end with Fable + ElevenLabs + Hedra.
Re-tune the prompts. The versions above are starting points. After 20 runs you'll know which constraints to tighten and which to drop. The prompt library that works for you is the one you've actually edited.
Free tier · From $10/month · Broadcast-quality lip-sync from any portrait
Frequently Asked Questions
8 questions answered
Claude Fable 5 is Anthropic's narrative- and language-craft model — tuned for tone, voice consistency, and prose that lands. It sits between Opus (peak reasoning) and Haiku (peak speed) and is the right default for any task where the writing matters more than the math. The API ID is claude-fable-5.
Opus 4.8 is the flagship — best at coding, agentic tool use, and dense reasoning. Fable 5 is roughly 3x faster and ~10x cheaper on output, optimised for matching voice references and producing fewer of the 'AI tells' that pattern-detectors catch. Use Opus when the answer has to be correct; use Fable when the answer has to be well-written.
Yes — paste them straight into either surface. For Cursor and Claude Code, switch the model selector to Fable 5 before running the prose-heavy prompts (DEV-2 PR writer, DEV-4 docs, DEV-6 error messages). Keep Opus 4.8 selected for the actual code edits and agentic loops.
Roughly $0.30 input / $1.50 output per million tokens, about a tenth of Opus 4.8's output pricing. The 90% prompt-cache discount applies the same way it does on Opus, so Voice Projects with large reference contexts stay cheap on repeat calls.
Yes — Anthropic's usage policy permits commercial use of model outputs on every paid tier, including for ads, YouTube monetisation, paid client deliverables, and shipped product copy. Check the per-tier ToS if you're in a regulated industry (legal, medical, financial).
Use Opus 4.8 for code that has to compile and pass tests, hard architectural reasoning, and any task involving math or formal logic. Use Fable 5 for video scripts, PR descriptions, commit messages, docs, error messages, marketing copy, and anything where voice consistency matters more than peak reasoning. Use Haiku 4.5 for high-volume classification and batch summarisation.
Build a Voice Project in claude.ai: drop in your 3–5 strongest pieces of approved writing (scripts, PRs, blog posts), a 1-paragraph brand-voice description, and 2 examples of writing you've explicitly rejected with one-line reasons. Run every prompt inside this Project. Fable picks up rhythm, vocabulary, and structural choices from the reference set without being told.
Both work. Many creators draft the outline (YT-3) in Fable, then use Fable for the hook and CTA (where voice matters most), and write the middle sections themselves to keep the human signal high. Others run end-to-end on Fable and edit in pass-two. The rule that helps: never ship a Fable draft you haven't read aloud once — your ear will catch the tells your eyes miss.