The most common complaint about AI writing tools is that the output "sounds like AI." In almost every case, the actual problem is the prompt — generic instructions produce generic output, regardless of which model is on the other end. We touched on this briefly in which AI tools actually save time; this is the specific technique.
Give It a Constraint, Not Just a Topic
"Write a tagline for my SaaS product" has no wrong answers, which means it also has no particularly right ones. A constraint — a length limit, a format, a thing to avoid — narrows the output toward something usable.
❌ Write a tagline for my SaaS product. ✅ Generate 10 tagline options (under 8 words) for [product], which does [description] for [audience]. Avoid generic SaaS tagline patterns like "[Verb] your [noun]. [Benefit]."
Why it works: The constraint does two things at once — it narrows the output space, and naming the cliché to avoid stops the model from defaulting straight to it.
Ask for Variations, Not a Final Answer
A single draft invites you to either accept something mediocre or start over. Multiple variations invite you to mix, compare, and edit — which is a fundamentally better editing posture.
❌ Write a subject line for this cold email. ✅ Generate 8 subject line options for this cold email. Make at least 3 curiosity-driven and 3 direct/specific.
Why it works: Explicitly requesting different approaches (not just different phrasings of the same approach) forces genuine variety instead of eight near-identical options.
Paste In Real Context, Not a Summary
Describing your product from memory loses detail the model could otherwise use. Pasting the actual text — a real customer quote, an actual feature list, a real competitor's homepage copy — gives it something concrete to react to instead of something to invent.
❌ Rewrite this feature bullet to sound more benefit-focused. ✅ Rewrite this feature bullet as a value statement using the "so what" test — state the outcome the customer gets, not what the feature does: [paste the actual bullet].
Why it works: The model can transform real input far more precisely than it can invent good output from a vague description of that input.
Name the Voice You Don't Want
"Sound professional" or "sound friendly" are directions with no shared definition — everyone's mental model of "professional" differs. Naming a specific failure mode to avoid is more effective than naming a vague tone to hit.
❌ Make this sound more human. ✅ Rewrite this so it doesn't sound like a press release — no "we're thrilled to announce," no exclamation points, no adjectives that aren't backed by a specific detail.
Why it works: "Human" is subjective. "No exclamation points, no unearned adjectives" is a rule the model can actually apply.
Ask It to Explain Its Choices When It Matters
For anything you'll actually publish (not just draft), asking the model to briefly justify each option gives you the information to choose well instead of guessing which one is strongest.
✅ Select and lightly edit the 3 strongest of these 5 testimonials for a landing page, prioritizing quotes with specific numbers or outcomes over general praise. Explain why you picked each one.
Why it works: The explanation surfaces the model's actual reasoning, which either confirms your instinct or reveals it picked for a reason you disagree with — both are useful.
Free Tool
🤖 AI Prompt Library
25 tested prompts for cold email, landing pages, positioning, and content — searchable, filterable, and copy-ready. Built from the same frameworks used across the site.
The Shared Pattern
Every rewrite above does the same thing: replace a vague instruction with a specific one. Constraint instead of open-ended. Variations instead of a single answer. Real text instead of a paraphrased description. A named failure mode instead of a vague tone. The prompt is doing the work most people expect the model to do on its own — and it's the difference between output you edit for ten minutes and output you rewrite from scratch.