Fable 5: The Prompt Cheat Sheet Just Expired

What Anthropic's newest model tells us about where the real skill now lives

In summary:

• Anthropic has published official guidance for Claude Fable 5, its newest and most capable model — and it retires much of what "prompt engineering" taught us.
• Detailed step-by-step prompts, written to compensate for weaker models, can now work against you: the model follows your rules even when the situation calls for something else.
• The new recommendation: state the goal, explain why it matters, provide background, set boundaries — in a word, context.
• Context is the one thing a model cannot generate for itself. It comes from your experience.

For the past three years, an entire cottage industry has grown around "prompt engineering". Magic phrases. Fifteen-rule templates. If you ever felt behind because you hadn't memorised the incantations, I have good news: the incantations are being retired — by the people who build the models.

Anthropic recently released Claude Fable 5, the first of its Claude 5 family, and published official guidance on how to work with it. The most interesting thing about that guidance is not what it adds. It's what it removes.

The old advice was to spell everything out: every step, every rule, every edge case. That made sense when models were weaker — detail compensated for limited judgement. Anthropic now warns that the habit has flipped from helpful to harmful: a newer model follows your fifteen steps faithfully even when the situation calls for something your steps didn't anticipate. The scaffolding that once held the model up now boxes it in. Their recommendation reads almost like a management course: state the goal, explain why it matters, provide the relevant background, define what the model should not do, and let it work out the how.

Notice what survives that list. The clever wording is gone. What remains — goal, why, background, boundaries — has a simpler name: context. And context is precisely what a model cannot generate for itself. It can reason brilliantly about the situation you describe; it cannot know that your Montreal office works differently from your Toronto one, that the Q3 number is misleading because of a one-off, or that your most important client reads formality as coldness. That knowledge lives in one place: your experience.

Anthropic's own documentation describes the model as something like a brilliant new employee who lacks context on your norms and workflows. If the model is the brilliant newcomer, who holds the advantage in the room? Not the person with the best template — the person who knows which details matter, and where the machine's suggestions should end and their own decisions begin.

Two caveats. This describes generative AI assistants specifically, not AI as a whole — a distinction I insist on. And details here are current as of July 2026; the direction of travel is what counts, and it points one way: away from learning the machine's language, towards knowing your own work deeply enough to explain it.

If that second skill sounds like yours, you were never behind. The industry just caught up to where you already were.