What I Changed in My Prompts for Claude Fable 5

Claude Fable 5 behaves differently enough from previous models that your existing prompts and habits will work against you. If you’re migrating, expect three things upfront: longer turns, stronger instruction following, and a model that actually performs better on harder tasks.

Here’s what I’ve learned about working with it.

Effort controls everything

The effort parameter is the single biggest lever. Use high by default, xhigh for critical work, and medium or low for routine tasks. The surprising part: lower effort on Fable 5 still outperforms xhigh on older models, so you can save cost on simple work without a quality penalty.

Rein in scope creep

Higher effort comes with a side effect: the model does more than you asked. If you don’t want it refactoring, adding abstractions, or “cleaning up” surrounding code, say so explicitly. It will if you don’t.

Audit long autonomous runs

On long-running tasks, instruct the model to verify every progress claim against actual tool results before reporting. This one instruction eliminates fabricated status updates, where the model reports completing work it never actually did.

Set hard boundaries

Define both what the model should do and what it should not. Without explicit boundaries, Fable 5 takes initiative in ways you didn’t ask for: drafting emails, creating backup branches, and other “helpful” extras nobody requested.

Give it memory

A simple Markdown file where the model logs one lesson per entry dramatically improves performance across sessions. It’s a small setup cost for a model that stops repeating its own mistakes.

Use parallel subagents

Parallel subagents finally work well. Use them aggressively, with three guidelines: keep them async so they don’t block the main thread of work, keep them long-lived rather than spawning fresh ones constantly, and make them cache-aware to control cost.

Migrate your prompts carefully

This is the trap most people fall into. Skills and system prompts written for older models tend to be too prescriptive, and on Fable 5 they degrade output instead of improving it. Strip them back and let the model’s defaults handle more than you’re used to.

One specific gotcha: any instruction that tells the model to echo or expose its reasoning will trigger the reasoning_extraction refusal classifier. Remove those instructions from migrated prompts.

Resources

• Prompting Claude Fable 5 — official guide