Most people walk up to AI like it’s a vending machine. Punch in one line, expect a result. When the result is bland, they blame the machine.

The problem isn’t the machine. You didn’t tell it what you wanted.

Here are 5 mistakes that kill your prompts and 6 techniques that fix them.

5 mistakes that kill your prompts

1. You’re googling, not briefing

You type “write a marketing email” and expect something usable. But the model isn’t pulling a pre-written answer from a database. It’s building something new, from scratch, based entirely on what you gave it.

You gave it 6 words. So you got 6 words worth of thinking.

2. Your prompt is a wall of text

Instructions, background, examples, formatting rules, all jammed into one paragraph. The model has to guess what’s important. When it guesses, it averages everything out and the output goes flat.

The thing you actually cared about ends up buried under a sentence you tossed in as an afterthought.

3. You didn’t say who it’s writing for

No role. No audience. No voice.

The model defaults to writing for everyone, which means it writes for no one. That generic, could-be-anyone tone people complain about? That’s what happens when you skip this step.

“You’re a copywriter” gets you halfway. “You’re a conversion copywriter for mobile apps targeting busy parents” gets you the rest of the way.

4. You gave up after the first response

First drafts aren’t final drafts. Not for humans, not for AI.

The first draft is AI slop. Iteration is what turns it into writing. When you see bad AI writing in the wild, you’re looking at a first draft that nobody bothered to push back on.

5. You fed it scraps instead of the full picture

Modern models can hold massive amounts of context. Google’s Gemini 2.5 Pro processes up to 1 million tokens in a single request, roughly 1,500 pages of text. Most people paste in a paragraph summary.

It can’t read your mind about everything you didn’t include. Give it the full picture:

  • Full documents
  • Full transcripts
  • Full briefs
  • Past versions of the thing you’re rewriting
  • The reference material you’d hand a contractor

Then ask the model to ask you questions back. End your prompt with something like “What am I not thinking of? Ask me about my blind spots. Be my red team.” Most people never do this. The model spots gaps you’d never catch on your own.

3 techniques that fix most bad prompts

1. The briefing framework

4 parts. Role, Context, Task, Format.

You’re not asking a favor. You’re assigning work. Think of it like handing a new contractor their first project. They don’t know your company, your audience, or your preferences. Everything they need has to be in the brief.

Copy this template:

[ROLE]
You're a [specific role] writing for [specific audience].

[CONTEXT]
Background: [what the model needs to know about the situation]
Constraints: [word count, tone, things to avoid, things to include]

[TASK]
Write a [specific deliverable]. It should [do this specific thing] for [this specific reader].

[FORMAT]
Structure it as: [exactly how you want the output organized]

Here’s the same template filled in:

[ROLE]
You're a retention marketer for a B2B project management tool.

[CONTEXT]
Background: Users who signed up for a free trial but haven't logged in for 7 days.
The tool's main selling point is automated deadline tracking.
Constraints: Under 120 words. No corporate jargon. No fake urgency.

[TASK]
Write a re-engagement email that reminds them of the one feature
most likely to pull them back in.

[FORMAT]
Subject line
Preview text (under 50 characters)
Body (3 short paragraphs max)
One CTA button text

The difference is the second prompt is a clear task. The model stops guessing and starts working.

2. The one example rule

Stop describing your writing style with contradictory adjectives.

“Friendly but professional.” “Concise but detailed.” “Casual but authoritative.” You just started a fist fight in your prompt. The model splits the difference and you get mush.

Instead, paste one paragraph of your own writing. One sample beats a page of style instructions.

Copy this template:

Here's how I write. Match this voice exactly:

"[Paste one paragraph of your writing here. A blog intro,
an email you liked, a social post that sounded like you.
Doesn't matter what it's about. The model will pick up the
rhythm, word choices, and sentence length.]"

Now write [your new topic] in this same style.
Keep the same sentence length, same level of directness,
same personality.

You don’t have to describe your voice. Just show it.

3. Show what you don’t want

Most people only tell the model what to do. Telling it what to avoid is just as powerful. Maybe more.

“No buzzwords” kills the “in today’s rapidly evolving landscape” opener.

“No filler sentences” cuts the throat-clearing.

“Don’t open with a question” stops the default pattern you’ve seen a hundred times.

Every constraint you add removes one more AI default.

Copy this template:

Write a [content type] about [topic] for [audience].

Rules:
- No buzzwords or corporate jargon
- No sentences that start with "In today's..."
- No filler transitions like "Moreover" or "Furthermore"
- Under [X] words
- Use real numbers or examples, not vague claims
- Don't open with a rhetorical question
- End with [specific element: a recommendation, a question, a next step]

[Add your own rules. What do you always have to edit out
of AI writing? Put it here. That's your constraint list.]

Think of constraints as a filter. Without them, the model picks from everything it knows about how that content type usually sounds. With them, you’ve removed the generic options. What’s left sounds more like a person wrote it.

3 techniques most prompting guides skip

1. Tell the model why

“Don’t use jargon” is a fine rule.

“Don’t use jargon because this is for first-time founders who won’t know the terms” is a better one.

When the model understands the reason behind a rule, it applies the rule in situations you didn’t specifically cover. It’s the difference between following the letter of the instruction and understanding the spirit.

Copy this template:

Write a [content type] about [topic].

Rules:
- No jargon, because [who this is for and why jargon fails them]
- Keep paragraphs under 3 sentences, because [where this will be
  read, like mobile or email, and why short blocks matter there]
- Use second person ("you"), because [what feeling you want the
  reader to have and why direct address creates it]
- Include one specific example per section, because [why your
  audience needs proof, not just claims]

Compare that to the same rules without reasons. The “because” version doesn’t just tell the model what to do. It tells the model what you’re trying to achieve. A model that understands the goal makes better decisions on the stuff you didn’t think to specify.

2. Draft, critique, revise

1 model, 2 roles. Writer and editor.

Most people stop at the first response and either accept it or start over. There’s a step in between that changes everything.

Prompt 1, the draft. Use the briefing framework from earlier:

[ROLE]
You're a [specific role] writing for [specific audience].

[CONTEXT]
Background: [what the model needs to know]
Constraints: [word count, tone, things to avoid]

[TASK]
Write a [specific deliverable] that [does this specific thing].

[FORMAT]
[How you want the output organized]

Prompt 2, the critique:

Look at what you just wrote. Find the weak spots.

- Where are the vague claims that need specific examples?
- Where does it sound like generic AI writing?
- What would a skeptical reader push back on?
- Are there any sentences that don't earn their spot?

Be honest. I'd rather fix problems now than publish something mediocre.

Prompt 3, the rewrite:

Rewrite the draft using your own critique.

Fix the weak spots you found. Cut anything that doesn't earn its place.
Replace vague claims with specific examples. If a sentence sounds like
generic AI writing, rewrite it in plain language.

Keep what worked. Tighten the rest. Don't start from scratch.

Output the full revised draft, not just the changed parts.

The first draft gets you to 70%. The critique finds what’s wrong. The rewrite gets you to 90%+. Total time for all 3 steps is usually under 5 minutes, less than most people spend manually editing a bad first draft.

One last thing

Most bad AI output traces back to one thing. The prompt didn’t give the model enough to work with. Fix that and most other problems disappear.

The shortlist:

  • The briefing framework: Role, Context, Task, Format
  • One example beats a page of style instructions
  • Tell the model what to avoid, not just what to do
  • Add “because” to your rules so the model understands the goal
  • Draft, critique, revise instead of stopping at the first response
  • Ask the model to ask you questions

The best prompts don’t ask for magic. They give clear assignments.

If you want more breakdowns like this in your inbox each week, here’s how Prompt Teardown works.