a beginner's guide
Anthropic's most capable model rewards a different skill than the one you've been practicing. Much of what you learned about prompting is backwards now, and the fix is mostly deletion.
Stop programming the route. Specify the destination.
tap the button — watch a prompt lose weight →
the same prompt, before and after
You are a world-class senior analyst with 20 years of experience.
Think step by step and show all your reasoning.
Step 1: pull the churn data. Step 2: list anomalies. Step 3: rank by severity. Step 4: summarize.
Format your answer as exactly: Summary (3 bullets), then a table, then a conclusion.
Look at last quarter's churn and tell me what's actually going on.
Every line above except one is doing the model's job for it.
With earlier models, good prompting meant more instruction: step-by-step procedures, output checklists, role-play framing, “think step by step.” Those instructions filled in what the model couldn't figure out on its own.
Fable can. Anthropic's own guide says prompts and skills written for older models are often too prescriptive; the scaffolding that helped before now gets in the way. One short principle replaces a checklist of ten rules. The skill shifts to naming what the goal is, why it matters, how you'll judge success, and where the boundaries are. Fable fills in the route.
tap any to expand — the navy lines are copy-paste
If you'd trust a smaller model with it, you're underusing this one.
Fable's advantage compounds with difficulty. Teams that only tested it on simple workloads underestimated it. Give it the thing you've been avoiding because it felt too big.
It does better when it knows what you're trying to do.
Context steers the work. The official docs give this pattern:
“I'm working on [larger task] for [who it's for]. They need [what the output enables]. With that in mind: [request]”
State the goal and success criteria. Don't dictate steps.
The step-by-step you'd write for a weaker model makes Fable's output worse. It finds the route better than you'd describe it.
Tell it how you'll judge the result.
One sentence of judgment beats three paragraphs of format spec:
“Good enough that I could hand it to my accountant without explaining anything.”
Capable models act when you only wanted a diagnosis.
If you're describing a problem or thinking out loud, say so. One boundary sentence does it:
“The deliverable is your assessment. Report findings and stop. Don't fix anything until I ask.”
Ambiguous, multi-threaded request? That's its job now.
For big tasks, make scoping the first deliverable:
“Scope this, ask me clarifying questions, then execute.”
Evidence-grounded progress, not guesswork.
In Anthropic's testing, this line nearly eliminated fabricated status reports on long-running work:
“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.”
Checking its own work cold beats self-congratulation.
Make it re-review 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.”
Don't pay for the tour of options it won't pick.
Left alone at high effort, Fable lays out every option. Steer it:
“If you're weighing a choice, give a recommendation, not an exhaustive survey.”
“Think step by step” is now a trap.
Asking for its hidden reasoning can trip Fable's reasoning-extraction safeguard, and your request gets answered by Opus 4.8 instead. Ask for the conclusion and the evidence behind it, not a transcript of how it got there.
Clarity, context, and concrete examples still win. What flipped is the procedural micro-instruction, the turn-by-turn directions weaker models needed. The basics didn't change.
take it with you
One page: the inversion at a glance, plus every copy-paste line — pin it next to your keyboard. Bundled with it: the Fable Prompt Builder skill — drop it into Claude, describe your task, and it assembles the Fable-ready prompt for you.