How to Write Better AI Prompts: 10 Rules

Picture this: you're a freelance grant writer with a stack of five proposals due this month, and you've just spent twenty minutes wrestling with an AI that keeps giving you the same bland paragraph no matter how you rephrase the request. That frustration is almost always fixable, and figuring out how to write better AI prompts is usually less about the tool and more about a handful of specific, learnable habits.

Before: The Weak Prompt

Write a paragraph about the impact of our after-school program on kids we can use in a grant proposal.

What this does: gives the model a topic but no data, no specific outcomes, no audience, and no format, so it fills the gaps with generic, forgettable language about "making a difference" and "empowering youth" that could describe literally any after-school program in the country.

⚠️ Common mistake: Treating a prompt like a search query instead of a briefing. The model isn't looking something up — it's generating text based on exactly what you give it, so vague input reliably produces vague output.

Why It Fails

This prompt fails for a specific, fixable reason: it's missing almost every ingredient a strong prompt needs. No data to ground the claims. No audience defined (a corporate foundation reads differently than a family foundation). No format constraint. No tone guidance. Each missing ingredient gives the model one more place to default to generic, average language instead of something specific to this particular program.

⚡ Pro tip: When AI output feels generic, check what's missing before rewriting from scratch — usually it's one of: specific data, a named audience, or a format constraint.

Here's a quick way to spot which ingredient is missing: read the AI's output and ask whether it could apply, almost word for word, to a completely different organization doing a completely different thing. If a rewrite for a food bank or an animal shelter would require changing almost nothing, the prompt was missing specificity somewhere. Generic output is really just a mirror reflecting a generic prompt back at you — the fix is never a "smarter" AI tool, it's a more specific ask.

After: The Improved Prompt

Role: You are a grant writer with experience in youth development nonprofits.
Audience: A corporate foundation that funds STEM education programs, evaluating dozens of similar proposals.
Data: Our after-school program served 140 kids last year; 78% improved their math grade by at least one letter grade; average attendance was 3.2 days per week.
Task: Write one paragraph (under 100 words) describing our program's impact, using the specific data above rather than general claims.
Tone: Confident and specific, not sentimental.

What this does: replaces every generic gap from the weak version with a specific fact or constraint, giving the model exactly what it needs to produce a paragraph a funder would actually find credible instead of one they've read a hundred times before.

⚡ Pro tip: Always feed the model your actual numbers. A model that has real data to cite will almost never default to vague language, because specific numbers are inherently more interesting to generate around than platitudes.

Breaking Down Each Element: 10 Rules for How to Write Better AI Prompts

The improved prompt above demonstrates several of the ten habits worth building into every prompt you write. None of these are complicated on their own — the value comes from applying them consistently instead of only remembering one or two when you happen to think of it:

1. State a role when expertise matters.
2. Name your specific audience.
3. Feed the model real data instead of asking it to invent examples.
4. Set a word or length constraint.
5. Specify tone explicitly rather than hoping the model guesses right.
6. Break multi-part tasks into numbered steps.
7. Give one example of your existing voice or a similar successful piece when consistency matters.
8. State what NOT to do when a specific failure mode keeps recurring, like "don't use exclamation points" or "don't include pricing."
9. Ask the model to show its reasoning for anything involving comparison, calculation, or judgment.
10. Treat your first version as a draft — test, then revise based on what specifically went wrong.

A freelance journalist applies rules 1, 2, and 9 almost every time she uses AI to draft interview question lists — role, audience, and reasoning shown before finalizing — because getting the order and framing of interview questions wrong wastes an actual limited window of a source's time, unlike a quick internal draft she can freely revise later. She's found that the reasoning step in particular catches questions that sound fine in isolation but would land poorly given who she's actually interviewing that week, whether that's a nervous first-time source or a seasoned executive used to deflecting pointed questions with practiced ease.

⚠️ Common mistake: Applying all ten rules to every single prompt, even trivial ones. A quick one-off question doesn't need a role, audience, and tone constraint — save the full structure for tasks where quality actually matters and you'll reuse the output.

A marketing coordinator at a nonprofit applied rules 3 and 4 alone — real data plus a length limit — to fix nearly all of her donor email drafts, without needing the full ten-rule structure every single time she sat down to write.

Variations for Different Contexts

Not every rule matters equally in every context. A quick internal Slack summary might only need rules 4 and 6 — length and structure. A client-facing proposal probably needs all ten. Match the rigor of your prompt to the stakes of the output; over-engineering a throwaway prompt wastes time, and under-engineering an important one wastes quality you can't easily get back.

It helps to think about the cost of a bad output on each side of that spectrum. A slightly-off internal Slack recap costs you thirty seconds of rereading. A slightly-off client proposal can cost you the client's confidence, or worse, get sent out before anyone catches the problem. Weighing that real cost, not just how "important" a task feels in the abstract, is usually the fastest way to decide how many of the ten rules a given prompt actually deserves.

⚡ Pro tip: Keep a personal "tier list" of your own recurring tasks — which ones need the full ten-rule treatment, and which ones are fine with a quick two-sentence prompt. This saves real time once you've used a task type a few times.

A communications director at a regional nonprofit built exactly this kind of tier list after a few months of trial and error. Board updates and funder reports landed in her "full ten rules" tier, since accuracy and tone both mattered enormously and mistakes were costly to fix after the fact. Internal meeting recaps and quick Slack summaries landed in her "two rules max" tier — just length and structure, nothing more — because nobody was scrutinizing word choice on a Tuesday status update. Sorting her own tasks this way up front saved her from over-thinking prompts that never needed it.

Save and Reuse This

Once you've built a prompt using several of these rules for a task you do regularly, that prompt becomes a template, not a one-time fix.

⚡ Pro tip: Number your saved templates by task type rather than by date. You'll find "grant proposal template" far faster than "prompt from March 14th" six months from now. Save it in PromptABCD so the next grant proposal, donor email, or client deliverable starts from something that already works instead of a blank cursor and the same twenty minutes of trial and error you started with the very first time.

The real payoff of treating these ten rules as habits rather than a one-time checklist shows up over months, not on the first prompt you write. Writers who internalize rules 1 through 4 specifically tend to find they barely need to think about them consciously after a while — role, audience, data, and format become automatic first questions for any new writing task, technical or not, AI-assisted or not. That's really the mark of having actually learned the skill, as opposed to just following a checklist: the questions stop feeling like a checklist and start feeling like the obvious first thing to ask about any piece of writing, the same way an experienced editor automatically checks for a clear thesis before worrying about word choice. Give yourself a few weeks of consciously running through the list before expecting it to feel automatic — most habits worth having take longer to stick than the first attempt suggests, and prompt writing is no exception to that general rule about building any new skill, technical or otherwise. Give it real time before judging whether the approach works for you. Most people who stick with it for a month never go back to writing prompts the old vague way again — the improvement in output quality is simply too noticeable to ignore once you've experienced it firsthand, on a real deadline, with real stakes attached to getting it right the first time, not just in a low-pressure practice exercise.