Stop prompting. Start briefing.

If you hand a competent freelancer one sentence of instruction, you get back something competent and generic. That isn't a failure of their ability. It's a failure of the brief.

The same thing happens with models, and it accounts for most of the gap between people who find these tools transformative and people who find them mildly useful. The tools are the same. The briefs are not.

What a prompt usually leaves out

A typical request — "write a summary of this document" — silently omits everything that determines whether the result is good:

• who is reading it and what they already know;
• what decision it needs to support;
• what must survive the compression and what can be dropped;
• how long it should be;
• what a bad version would look like.

A person would ask. A model won't, unless you invite it to — so it fills the gaps with the most average plausible answer. That's what "generic AI output" actually is: the average of the space you left undefined.

The four-part brief

The structure I keep returning to:

1. Role and context. Not "you are an expert" theatre — actual situation. "This goes to a finance lead who hasn't been in the project meetings."
2. The job to be done. The decision or outcome it serves, not the artefact. "They need to decide whether to approve another quarter of funding."
3. Constraints. Length, format, tone, what to exclude. Constraints do more work than adjectives.
4. What good looks like. One example, or a description of the failure mode you want avoided. "Don't restate the timeline — they have it. I want the two risks that would change the decision."

That's four sentences. It routinely makes the difference between output you rewrite and output you send.

Give it permission to push back

The single highest-return addition to any brief:

If anything here is ambiguous or you think I'm asking for the wrong thing, ask before you start.

Models default to compliance. They will happily produce a polished answer to a badly framed question. Explicitly inviting the challenge converts that silence into the most valuable part of the exchange — and it's usually where you find out your own thinking was fuzzy.

Iterate on the brief, not the output

The common failure pattern is to get a mediocre result and start editing it. You end up doing the work by hand, one correction at a time, and you learn nothing reusable.

The better move is to treat a bad result as a diagnostic. Ask what in the brief produced it. Then fix that and regenerate. Two or three rounds of this and you have a brief you can reuse for the next twenty documents of the same kind.

A brief you can steal

Context: <who this is for, what they already know>
Goal:    <the decision or outcome this supports>
Do:      <format, length, must-include>
Don't:   <the specific failure mode you keep seeing>
Check:   Ask me anything ambiguous before you start.

Keep it in a note. Fill in five blanks. It takes about ninety seconds and it is, in my experience, the highest-leverage ninety seconds available in daily AI use.

Why this keeps mattering

As models get better at reasoning, they get better at executing an underspecified brief in a plausible direction — which means the cost of a vague ask goes up, not down. The work of saying clearly what you want isn't a temporary workaround for a limitation. It's the durable part of the skill.