How to Write Better AI Prompts

Vague prompts produce vague answers. Specific prompts produce useful answers.

• Weak: "Write a tweet about our new feature."
• Strong: "Write a 220-character tweet announcing our new pricing page. Tone: casual but professional. Include exactly one emoji. No hashtags."

The strong version names the format, length, tone, and constraints. The AI doesn't have to guess.

The expected difference in output is dramatic. The weak prompt typically returns something like "Exciting news! We've just launched an amazing new feature... #innovation #tech" - Generic, hashtag-littered, unusable. The strong prompt returns copy you could publish with one small edit. Common mistake here: people specify the topic but not the constraints. Length, tone, and "do not do X" rules are where most of the quality lives.

Tell the AI who you are, who the audience is, and what's already true. Front-load the context.

• Weak: "Write a sales email."
• Strong: "I'm a B2B SaaS founder writing to existing customers who haven't logged in for 30 days. Goal: re-engage them with our new dashboard. Tone: friendly, no pressure. Length: 80-120 words."

A useful mental test: could a freelancer who's never met you complete the task from your prompt alone? If they'd have to email you three clarifying questions, the AI needs those answers too - It just won't ask. It'll guess, and its guesses are bland. Mention the audience's knowledge level ("they're non-technical"), what's been tried before ("we already sent one reminder, no response"), and the action you want ("goal: they click the dashboard link").

Ask explicitly for a numbered list, a markdown table, JSON, or 3 paragraphs. The AI will deliver a clean format you can copy-paste somewhere useful.

Example: "Return your answer as a markdown table with columns: Idea, Effort (1-5), Impact (1-5), Why. Sort by Impact descending."

Format instructions also quietly improve the content. Forcing an Effort and Impact score makes the model actually evaluate each idea instead of listing whatever comes to mind. Other formats worth stealing: "answer in exactly 3 sentences" (kills rambling), "give the answer first, then the reasoning" (stops the essay-before-the-point habit), and "output valid JSON only, no commentary" when feeding results into code.

One good example of the desired output is worth a paragraph of instructions. This is called few-shot prompting (as opposed to zero-shot, where the model gets instructions only).

Example: "Convert dates from US format to ISO 8601.
Input: Jan 5, 2024 → Output: 2024-01-05
Input: Dec 31, 2025 → Output: 2025-12-31
Now convert: Input: March 8, 2026."

This matters most for fuzzy, taste-based tasks. Want product descriptions in your brand's style? Paste two real ones and say "match this style." That beats any adjective list - "witty but warm" means nothing precise to a model, but two examples of witty-but-warm define it exactly. Common mistake: giving an example that contains quirks you don't want copied. The model imitates everything, including the flaws.

Don't expect the first answer to be perfect. Tell the AI what to fix:

• "Shorter and punchier."
• "More technical."
• "Add a counter-argument."
• "Less hedging - Pick a position."
• "Rewrite in plain English, no jargon."

The AI holds the conversation in memory, so it iterates much better than starting over each time. Two iteration habits separate power users from beginners. First, change one thing at a time - "shorter, funnier, and more formal" pulls in three directions and produces mush. Second, when an output is close, say what to keep: "Keep the first two paragraphs exactly as they are; rewrite only the ending." Without that, the model happily rewrites the parts you loved.

Telling the model who to be changes which patterns it draws on. "You are a contracts attorney reviewing this clause for a small-business client" produces meaningfully different analysis than a bare "review this clause" - More caveats where they belong, terminology used correctly, risks framed for the right reader.

For repeat use, set the role once as a system prompt so every message in the conversation inherits it. AskAI.free's specialised pages like the AI lawyer and AI accountant are exactly this: a strong role prompt baked in, so you skip the setup.

Don't overdo it. "You are the world's greatest genius marketer" adds nothing. Roles work when they're specific ("a paediatric nurse explaining to a worried parent"), not when they're flattery.

If you find yourself rewriting the same prompt structure repeatedly, save it. A good template has blanks for the parts that change and locked-in wording for the parts that work:

"I'm a [role] writing to [audience]. Write a [format] about [topic]. Tone: [tone]. Length: [length]. Must include: [requirements]. Never: [exclusions]."

Browse our library of 31 curated prompt templates for writing, coding, thinking, learning and more - Or save your own with the Pro plan.

Watch the rules combine. Dan starts with: "Write a job posting for a developer." The output is generic boilerplate - "rockstar developer," "fast-paced environment," 400 words of nothing.

Move 1, add context and specifics: "We're a 6-person fintech startup hiring our first senior backend engineer (Python/Postgres). Salary $140-160K, remote-first, equity included. Write a job posting." Better - Real details now appear.

Move 2, add format and constraints: "Structure: 2-sentence hook, What you'll do (5 bullets), What we need (4 bullets), What we offer (3 bullets). No clichés like 'rockstar' or 'fast-paced'. Under 250 words."

Move 3, iterate: "The hook is boring. Rewrite it around the fact that this hire owns the entire payments pipeline." Three moves, four minutes, and the posting is genuinely publishable. That's the loop for every prompt: context, constraints, iterate.