What is Claude Fable 5?

Claude Fable 5 is Anthropic's first publicly available Mythos-class model. It is built for long autonomous runs in harnesses like Claude Code: it plans, executes, checks its own progress, and refines work over hours or days without you babysitting every step.

Do these prompts work in regular Claude chat?

They work best in Claude Code or any setup where the model can act on files, run commands, and loop on a task. In one-shot chat you still get sharper output, but you will not get the overnight autonomous run Fable 5 is built for.

What should I hand off to Fable 5 first?

Pick the project you keep not finishing because the middle is boring: the migration, the rebuild, the tool you sketched and abandoned. Run the 2-hour proof prompt first. If the proof passes your morning review, scale to the overnight handoff.

8 prompts to delegate real work to claude fable 5.

Q: Does the July 7 window in this post's title still apply?

No. That was Fable 5's original access window on Pro, Max, Team, and Enterprise plans, and it has passed. Check Anthropic's current pricing page for what your plan includes today. The eight prompts below teach delegation, which works the same regardless of which plan or window you're on.

elisabeth hitz · july 6, 2026 · 7 min read

updated july 26, 2026: this post originally framed july 7 as an access deadline for fable 5. that date has passed, and anthropic's pricing and plan inclusion may have changed since this was published, so check anthropic's current pricing page for what your plan includes today. the eight prompts below are unaffected. delegation works the same no matter which plan or window you're on.

anthropic shipped claude fable 5, the first public mythos-class model. the headline is benchmarks. the actual shift is autonomy: it can run for hours or days in claude code while you sleep. these eight prompts teach the skill that matters more than better chatting: delegation.

the solo founder bottleneck was never intelligence. it was supervision. you write a prompt, babysit the output, correct it, write the next prompt. fable 5 breaks that loop if you scope the handoff cleanly. waste it as a faster chatbot and you learn nothing.

copy each block below into claude code. swap the bracketed parts for your project. run prompt 8 first if you have never handed off overnight work.

how to use this list

prompt 8 (2-hour proof) on a small slice of the real project.
prompts 1–4 to lock scope, constraints, done criteria, and checkpoints.
prompt 5 (overnight handoff) when the proof passes your morning review.
prompts 6–7 every morning until the project ships.

prompt 1 · definition of done

Before you write any code, define done for this project in a short checklist I can verify without reading your reasoning. Project: [name] Goal in one sentence: [outcome] Must ship: [files, pages, or artifacts] Must pass: [tests, lint, manual checks] Must not change: [paths, APIs, billing, prod data] Output only the checklist. No implementation yet.

prompt 2 · do-not-touch lock

You are working in this repo. List every path, env var, and external integration you must NOT modify without asking me first. Then propose a safe working branch name and a rollback plan if this run goes wrong (one paragraph, concrete commands). Repo context: [stack + main folders] Sensitive areas I already know: [auth, payments, prod config, etc.]

prompt 3 · deliverable manifest

For project [name], produce a deliverable manifest: every file you will create or edit, with one line on what each file must contain when finished. Group by: new files, edited files, deleted files (if any). End with a 5-bullet morning review script I can run in under 10 minutes.

prompt 4 · checkpoint cadence

Plan this as an autonomous run with self-checkpoints every [90] minutes. At each checkpoint you must: (1) restate the goal, (2) list what finished since the last checkpoint, (3) list blockers, (4) say whether you are still on track for done criteria. If you drift, stop and write a short recovery plan before continuing. Do not silently expand scope.

prompt 5 · overnight handoff

You are my overnight contractor for project [name]. Goal: [one sentence outcome] Constraints: [stack, style, deadlines, do-not-touch list from earlier] Definition of done: [paste checklist from prompt 1] Work autonomously until done criteria pass or you hit a blocker that needs my decision. When blocked, leave a single QUESTIONS.md with numbered options (recommend one). When finished, write HANDOFF.md: what changed, how to verify, what I should review first in the morning.

prompt 6 · abandoned project rescue

This project stalled because the middle was boring: [migration / rebuild / refactor / feature name]. Read the repo. Summarize current state in 8 bullets (what works, what is half-done, what is broken). Propose the smallest slice we can finish in one autonomous run that still moves the business outcome: [outcome]. Do not restart from scratch unless you justify why in 3 sentences.

prompt 7 · morning contractor review

I am reviewing your overnight work like a contractor deliverable, not pair programming. Read HANDOFF.md and the diff. Answer only: 1) Did you meet the definition of done? (yes/no per item) 2) Top 3 risks I should fix before shipping 3) One sentence: ship, iterate, or rollback Be blunt. No praise.

prompt 8 · two-hour proof run

Before any overnight run, execute a 2-hour proof on the hardest slice of [project name]. Pick the slice where failure would teach us the most. Timebox to 2 hours of autonomous work. Deliver: small working increment + HANDOFF.md + updated done checklist with what proof validated. If proof fails, stop and report why overnight delegation would be reckless.

what not to waste fable 5 on

one-shot captions, quick emails, and "make this sound better" tasks. that is chat work. fable 5 pays off on the project you keep not finishing because it is too big to sit through.

anthropic also routes some sensitive domains (cybersecurity, biology, chemistry, health) to opus 4.8 automatically. if answers feel slightly less sharp in those areas, you may have been handed off. that is expected.

after july 7

access economics may change when fable 5 moves off included plan usage. the prompts still work on any capable model in claude code. the durable asset is the delegation muscle: scoping, checkpoints, morning review.

if you want those rules pinned so every claude session starts on-spec (not drifting back to generic), that is what skill files are for. one free file streams live on the toolkit page. the full zip is the paid path on the same page.

is your prompt actually any good?

the free prompt scorer rates any prompt across five dimensions and tells you exactly what to fix. paste it in, see your score, no email wall.

score my prompt free

browse all 30 free tools →

sources: anthropic fable 5 launch materials (mythos tier, autonomous harness behavior, plan inclusion window), anthropic ai fluency framework (delegation as core skill). author tested overnight handoffs in claude code, july 2026.