Why Your AI Output Is Mediocre (It's the Brief)

Most complaints about AI output — generic, bland, missed the point — are briefing problems rather than model problems.

The test: would a competent freelancer who has never seen your business produce good work from what you just typed? Usually the honest answer is no.

The five-part brief

Not a magic formula. The same things you'd tell a person.

1. Context. Who you are, who this is for, and what it's part of.

We're a five-person bookkeeping firm serving trades businesses in the UK. This email goes to clients who've missed their month-end deadline twice.

2. Task. What you want, specifically.

Draft a short email asking for their receipts, firm but not annoyed.

3. Constraints. Length, tone, what to avoid, what must appear.

Under 120 words. Plain English, no jargon. Don't apologise for chasing. Must mention the 5th as the deadline.

4. Format. Structure, and how you want it delivered.

Subject line, then the body. Give me three subject line options.

5. An example. The highest-leverage part — see below.

Most people supply only the task. That's like handing a freelancer one sentence and being disappointed in the result.

One example beats three paragraphs of instruction

If you have something good, paste it.

Here's an email I sent that worked well: [paste]. Write one like this for a different situation: [describe].

An example carries tone, structure, vocabulary, length and register simultaneously — all the things that are tedious to describe and easy to demonstrate.

Two examples are better than one, because the model can infer what's consistent between them rather than copying incidental details from a single sample.

If you don't have a good example, an example of what you don't want also works:

Not like this: [paste]. It's too formal and it buries the request.

State what to avoid

Negative constraints are underused and frequently do more than positive ones, because they rule out the defaults you'd otherwise have to keep correcting.

Useful ones:

• No preamble. Start with the thing itself.
• No summary at the end restating what you just said.
• Don't hedge. State it plainly.
• No em dashes / no bullet points / no headings — whatever your house style excludes.
• Don't invent figures. If you don't have a number, leave a placeholder.

That last one matters more than it looks. It won't stop every fabrication, but it changes the default behaviour from filling gaps to flagging them.

Give it the raw material

The most common serious error: asking for something that requires information you haven't supplied.

"Summarise our Q3 performance" — with no data attached — produces invented plausible content.

If it needs a document, paste the document. If it needs figures, provide them. If it needs to know your pricing, tell it.

Anything you don't supply, it will either omit or invent, and invention is the dangerous mode because it looks like the real thing.

Iterate rather than restarting

When output is wrong, most people rewrite the whole prompt and try again. Correcting is faster and produces better results.

That's close. Make it shorter, drop the second paragraph entirely, and make the ask more direct — it currently sounds tentative.

Specific correction, not "make it better." You're editing, and the same rules apply as giving feedback to a person: observation, impact, request.

Three or four rounds of correction beats three attempts from scratch.

Reuse what works

Once a brief produces good output reliably, save it. Most small businesses re-type variations of the same request forever.

Keep them somewhere findable, with a placeholder for the variable part:

CLIENT CHASE EMAIL
Context: [firm description]
Task: Draft a chase email for a client who has missed [deadline] [n] times
Constraints: Under 120 words, plain English, firm not annoyed, no apology
Format: Subject line (3 options) then body
Example of good: [paste]

That's the same logic as any other documented process — write it down on the third repetition, and store it where the work happens.

What briefing can't fix

Worth being clear about, because better prompts get oversold as a solution to everything.

• It doesn't know your specific facts unless you supply them
• It doesn't know what's happened recently in your business or the world
• It can't make judgement calls that depend on context you haven't given it
• A confident wrong answer looks identical to a right one

Better briefs produce better output. They don't make the output verifiable — that's still on you, and it's why the verification test comes first.

The mistakes

1. Task only, no context. The single most common cause of generic output.
2. Describing tone instead of showing it. One example beats a paragraph.
3. No negative constraints. You keep correcting the same defaults.
4. Asking about material you haven't supplied. Produces invention.
5. Restarting instead of correcting. Slower and worse.
6. Never saving what worked. Re-typing the same brief forever.

What to do next

Take the last AI output that disappointed you and rewrite the brief with all five parts, including one example of what good looks like.

If the result is materially better — and it usually is — that's your answer about where the problem was.