Prompting Fable
The habits that worked on older AI can make Claude's most powerful model worse. Here's what to do instead — with the exact words to copy.
The whole thing in 30 seconds
Don't tell Claude every step. Tell it what you want and how you'll know it's good— then let it work out the steps.
- 1Give it your hardest problem, not your easiest.
- 2Say why you're doing this, not just what you want.
- 3Describe the result you want, not the steps to get there.
- 4Tell it how you'll know the result is good.
- 5Say whether you want it to just look, or actually make changes.
- 6Let it plan the work before it starts.
- 7Ask it to back up its progress with proof.
- 8Ask it to check its own work with fresh eyes.
- 9Ask for its pick, not a list of every option.
- 10Don't ask to see its private thinking — it can backfire.
The idea, in motion
Ten tips. Here's the whole idea first.
Don't hand it the steps. Hand it the goal— and how you'll know it's good.
Bring your hardest problem.
It gets better the harder the task. Say why you're doing it, then let it plan.
One line about “good” beats a page of rules.
Ask for proof. Ask for its pick, not a list.
And don't ask to see its hidden thinking — that one can backfire.
Say why. Show what good looks like. Then let it work.
An animated explainer of the same guide, in about 24 seconds.
The full guide
Don't hand it the steps. Hand it the goal.
With older AI, good prompting meant moreinstruction — long checklists, “do it step by step,” spelled-out rules. That filled in what weaker models couldn't do on their own.
This model can do it. Anthropic's own guide says prompts written for older models are often too bossynow — the extra hand-holding gets in the way. The new skill is naming the goal, why it matters, and how you'll know it's good. Then it finds the steps.
1Give it your hardest problem
If a smaller, cheaper AI could do it, you're wasting this one.
This model gets better the harder the task. Bring the thing you've been putting off because it felt too big.
2Say why, not just what
It does better work when it knows what you're trying to do.
Give it the reason. This is the pattern from the official guide:
I'm working on [larger task] for [who it's for]. They need [what the output enables]. With that in mind: [request]3Describe the result, not the steps
Say what a good result looks like. Don't list every step.
The step-by-step you'd write for a weaker model actually makes this one's work worse. It finds the path better than you can describe it.
4Tell it what "done" looks like
One line about "good" beats three paragraphs of rules.
Give it the test you'll judge it by:
Good enough that I could hand it to my accountant without explaining anything.5Say whether to look or to fix
Powerful AI will start fixing when you only wanted it to look.
If you're just thinking out loud, say so. One line does it:
The deliverable is your assessment. Report findings and stop. Don't fix anything until I ask.6Let it plan first
Big, messy, many-part task? Let it plan before it starts.
Make the plan the first thing it gives you:
Scope this, ask me clarifying questions, then execute.7Ask for proof
Get evidence-backed progress, not guesses.
In Anthropic's testing this line nearly stopped made-up status reports:
Before reporting progress, audit each claim against actual evidence. Only report work you can point to evidence for; if something isn't verified, say so.8Add a fresh-eyes check
It checks its own work better cold than while patting itself on the back.
Make it re-check its own finish line:
When you think you're done, verify your work against the success criteria as if you were a skeptical reviewer seeing it for the first time.9Ask for a pick, not a list
Don't pay for a tour of options it won't choose.
Left alone, it lists every option. Steer it:
If you're weighing a choice, give a recommendation, not an exhaustive survey.10Don't ask for its hidden thinking
"Think step by step" is now a trap.
Asking to see its private reasoning can trip a safeguard and hand your request to a different model. Ask for the answer and the proof behind it, not a play-by-play.
What didn't change
Being clear, giving context, and showing an example still win. What flipped is the step-by-step hand-holding weaker models needed. Also: longer waits are normal on hard tasks — it's working, not stuck.
Take it with you
Get the cheat sheet + the free skill
Run the skill in Opus or Sonnet and it writes a Fable-ready prompt for you — then you paste that into Fable 5. The cheat sheet is one page — print it, keep it by your screen.
No spam, ever.