Say the what, delegate the how: specify the outcome, not the recipe

Overview

Within weeks of each other in mid-2026, two frontier labs told their users the same surprising thing: delete instructions from your prompts. Anthropic’s guidance for Claude Fable 5 warns that prompts and skills written for older models are “often too prescriptive” and “can degrade output quality.” OpenAI’s prompt guide says the same of GPT-5.5: “legacy prompts often over-specify the process because earlier models needed more help staying on track.” After three years of “be more specific”, the newest advice is to specify less — of one particular kind.

By the end you’ll be able to tell apart the two things your instructions specify — the what and the how — and stop scripting the one that now costs you quality.

The content

The obvious read is that a better result comes from a more specific prompt: spell out every step, leave nothing to chance. The overturn is that specificity has two objects, and they’ve come apart. There’s the what — the goal, the constraints that genuinely matter, and what a finished, correct result looks like. And there’s the how — the method you’d use to get there: do this first, then that, structure it this way. Be as exact as ever about the what. Stop dictating the how.

Here’s why the split matters now. When you script the method, the model follows your script — even where it could have taken a better route. On older models that hand-holding was protective; without it they’d wander. On current ones it narrows the search space for no gain. OpenAI’s guide is specific about the damage: over-specifying “can add noise, narrow the model’s search space, or lead to overly mechanical answers.” Anthropic’s is blunter still — the instructions that helped last year “can degrade output quality” today. That old instinct to over-direct is now a tax.

So the move worth building is one habit: say the what, delegate the how. Developer Simon Willison, writing about coding agents on 3 July 2026, put it plainly — tell the model to “use its own judgement” rather than dictating how it works. The lab guidance generalises it well past code: OpenAI advises defining “the target outcome, success criteria, constraints, and available context, then let the model choose the path.” It’s the same direction Anthropic sees the whole field moving — “letting intelligent models act intelligently, with progressively less human curation.”

The trap — and this is the part to get right — is mistaking a real constraint for micromanagement. “Cite every figure”, “stay under 300 words”, “don’t contact the customer”, “use our house template”: those aren’t method, they’re guardrails, and they belong in the what. Keep them explicit. The skill is telling a guardrail apart from a habit of over-directing. And delegating the how raises the price of not checking: when you didn’t dictate the route, you owe the result a proper read — which is exactly where a plan-first gate pays, letting the model pick the route, then showing you the one it chose before it runs.

Try it

Take one of your standard multi-step prompts — the kind where you’ve written out the procedure. Rewrite it with the model using this:

Here is a prompt I normally use:
"""
[paste your step-by-step prompt]
"""

Split what it contains into two lists:
1. THE WHAT — the goal, the constraints I actually care about, and what a
   finished, correct result looks like.
2. THE HOW — the parts that only describe the method or the order of work.

Then give me back a new version that keeps list 1 in full, drops list 2,
and ends with: "Use your own judgement on how to approach this. Before you
start, show me your plan."

Run the new version on a real task and compare it to your old step-by-step one. Where it breaks: if the result drifts from what you needed, check whether a genuine requirement got sorted into “the how” — a constraint disguised as a step. Put it back as a guardrail, not a script. That sorting is the skill.

Additional reading

• Fable’s judgement — Simon Willison’s Weblog (2026-07-03) — the practitioner framing: tell the model to “use its own judgement” on method rather than dictating it. Written about coding agents; the principle carries.
• Prompting Claude Fable 5 — Claude Docs (living page, June 2026) — source of the warning that older prompts and skills are “often too prescriptive… can degrade output quality”, and of “steer most behaviors with a brief instruction rather than enumerating each behavior by name”; framed around documents, spreadsheets and analysis, not just code.
• Prompt guidance — OpenAI API docs (living page) — GPT-5.5 guidance: “define the outcome and leave room for the model to choose an efficient solution path”, and that “legacy prompts often over-specify the process because earlier models needed more help.” Newer models reward the outcome, not the procedure.
• Effective context engineering for AI agents — Anthropic (2025-09-29) — the trajectory this sits on: “agentic design will trend towards letting intelligent models act intelligently, with progressively less human curation.”

Editor’s note

Judgement is the word of the moment at this time in AI competence building — or at least it is for me. On this topic, judgement has to work overtime. This rule is not universal. You’re not carving up prompts and suddenly getting better results out of last year’s models. This is a July 2026 flagship model development, where your ability to define and articulate the goal has become a more productive thing than your ability to describe the steps you want a model to follow. But clearly, the judgement about what to leave in as a required step is more important now than ever. You can’t leave out the requirement to refer to a critical document, but you might be degrading your output quality if your instruction is overly prescriptive in relation to what to refer to within the context subset. This is another example of the advice changing at the rate of model releases. Personally, I think that it’s exciting, however challenging it may be.