How to write better AI prompts

Most AI prompts fail for the same reason: they tell the model what topic to cover, but not what to do.

"Write something about climate change" gets you a mediocre essay. "Write a 400-word op-ed arguing that personal carbon footprints are a distraction, written for a skeptical conservative reader, in the voice of a New York Times opinion columnist" gets you something usable.

The difference isn't AI magic. It's instruction quality. Here are the five moves that consistently raise output quality without making prompts longer than they need to be.

1. Start with a clear action verb

The first word of your prompt should be a verb that names the operation. Write. Generate. Draft. Summarize. Translate. Analyze. Compare. List. Critique. Each verb implies a different output shape.

Compare:

Climate change essay

vs.

Write a 400-word essay arguing climate change requires structural policy, not individual action.

The first prompt forces the model to guess your intent. The second one removes that guess.

A prompt with no action verb almost always produces hedged, generic output. The AI Prompt Counter flags this for you under "No clear action verb" — that single signal predicts a lot of mediocre output.

2. Name the audience

Models trained on the public internet have a default voice: middle-of-the-road, mildly enthusiastic, slightly hedged. To break out of it, name your reader.

Explain how a transformer model works.

becomes

Explain how a transformer model works to a senior backend engineer who has never worked with ML before. Skip the hype. Assume they understand vectors and gradient descent.

Now the model has to make audience-aware choices: vocabulary, what to skip, what to motivate.

The audience can be specific — "a Series B founder evaluating AI tools" — or a role — "a confused first-time user." Either way, it constrains the space.

3. Specify the output format

The shape of the output matters as much as the content. Specifying the format is the single highest-leverage edit you can make.

Examples:

• "Return as a JSON object with keys name, summary, risks."
• "Format as a markdown table with columns Tool / Best for / Free tier."
• "Three bullet points, each under 20 words."
• "A reply email no longer than four sentences."
• "Output only the code, no commentary."

Format specs prevent the most common failure mode: the model returning a wall of prose when you wanted a list, or a list when you wanted a paragraph.

4. Set constraints

Constraints sharpen output the way deadlines sharpen work.

• Length: "under 400 words", "exactly 3 paragraphs"
• Tone: "professional but not corporate", "warm but precise"
• Style: "in the voice of Paul Graham", "the way Stripe Press writes"
• Don't: "avoid jargon", "no hedging language", "no LinkedIn voice"

A useful rule: every constraint you add removes a class of failure modes. "No hedging" removes the "I think... it could be argued..." opener. "No bullet points" removes lazy list-formatting. "Under 200 words" removes filler.

5. Add one example when stakes are high

For style-sensitive tasks — copy, voice, tone — one example beats ten adjectives.

Write a product launch tweet for our new feature.

is weak. This is stronger:

Write a product launch tweet in this style — short, no hashtags, ends with a concrete benefit:

"We rebuilt the editor. It's twice as fast and three pixels prettier. Try it."

Now the model has to imitate, not invent. For copy and voice work, this is the highest-ROI move you can make.

Putting it together

Apply all five and you get something like:

Write a 250-word product launch announcement for a developer audience, posted to our blog. Voice: technical, plainspoken, slightly dry — like Stripe Press. Format as three short paragraphs followed by a one-line CTA. Don't use the word "excited" or marketing hype. Example tone:

"We shipped a thing. It does the thing well. Try it."

This is verbose compared to "write a launch post," but each addition removes a class of bad output. The total length is still under 80 words.

A test you can run today

Open the AI Prompt Counter and paste your current prompt. The Prompt Score panel will tell you which of these five moves are missing. Most prompts that score below 60 are missing three or more.

Make one move at a time. Add an action verb. Add an audience. Add a format spec. Re-run the model. The change in output is usually obvious within one or two tries.