The first month I used AI to draft blog posts, I deleted almost everything it gave me. The tool was capable. My prompts were the problem. I was typing things like “write a blog post about email marketing,” and what came back was technically an article: an intro that warmed up for three sentences, a few tidy subheads, a closing paragraph that restated the opening. It was also flat, over-hedged, and read like it had been pressed out of the top ten search results, because functionally that is what it was.
The change did not come from a better model. It came from learning to ask in a way that took away the model’s escape routes. A vague prompt lets it retreat to the safest, most average version of a topic, the version every other site already published. A precise prompt closes those exits and forces something with a point of view. Below are the patterns I keep reaching for, written out as the real prompts I use.
Why most blog prompts fail
Before the fixes, it helps to see why the default output feels the way it does. Almost every weak result traces back to one of these gaps. Give the model none of this context and it fills the space with the median of everything it has read.
• No reader in mind. “Write about X” names a topic, not an audience. The model hedges to cover everyone at once, so the writing lands with no one in particular.
• No angle. Without a claim to argue, it summarizes. Summaries are exactly what already ranks, and they give a reader no reason to keep going.
• No format. Left to choose, it defaults to intro, three vague headers, conclusion. Readers now recognize that shape on sight and bounce.
• No voice. With no sample to match, it falls back on its house style. That style is the thing people point at when they say a piece sounds like AI.
• One giant ask. Requesting a whole finished post in a single prompt means you cannot steer. You get one roll of the dice and an edit-from-scratch job.
The parts of a prompt that works
A prompt that produces usable copy tends to carry seven kinds of instruction. You do not need all seven every time, but the more you specify, the less the model is left to guess.
| Component | What it tells the model | Vague vs specific |
|---|---|---|
| Role | Who is writing and the stance they take | “a writer” vs “a skeptical B2B copywriter who has run these campaigns” |
| Reader | Who this is for and what they already know | “everyone” vs “ops managers at 50-person startups, technical but time-poor” |
| Intent | What the reader should do or believe after | unstated vs “leave convinced they can cut their tool stack in half” |
| Angle | The single claim the piece argues | “about pricing” vs “why usage-based pricing quietly punishes your best customers” |
| Format | Structure, length, and sections | unstated vs “1,100 words, a hook, three H2s, one table, a two-line takeaway” |
| Voice | Tone plus a real sample to match | “professional” vs a pasted paragraph of your own writing |
| Grounding | The facts it is allowed to use | unstated vs “only the figures in the notes below; flag anything unsupported” |
Six patterns, with the prompts
These are the moves that changed my output the most, in rough order of payoff. Each one is a small addition to the prompt, and each closes a specific escape route.
1. Name the reader and the stakes
The highest-return instruction is telling the model exactly who is reading and what they are trying to do. It changes the vocabulary, the examples, and how much the piece stops to explain.
Prompt
| Write a 1,200-word post on intermittent fasting for parents in their late 30s who have started and quit diets before. Angle: the meal-timing rules matter less than people think, and consistency is the only real predictor of results. Practical and a little skeptical. Assume the reader is smart and short on time. Use two concrete examples of daily eating windows. |
2. Give it a real voice sample, not an adjective
Telling the model to “sound human” or “be conversational” does almost nothing. Those words mean everything, and so they mean nothing the model can target. Paste two or three paragraphs you have actually written and ask it to match the rhythm and word choice. This is the closest thing to an off switch for the AI-voice problem.
Prompt
Here are three paragraphs I wrote. Match this sentence rhythm, vocabulary, and level of directness. Do not smooth it out or make it more formal: [paste your own writing] |
3. Constrain the shape before the words
Decide the structure yourself and hand it over. When you leave structure open, you get the default article skeleton every time. Name the parts and their limits, and the draft arrives in a form you can publish rather than a form you have to rebuild.
Prompt
| Structure: a one-sentence hook, a short first-person intro (no throat-clearing, no “in today’s world”), three H2 sections, a comparison table, and a two-line takeaway. Keep paragraphs to four sentences or fewer. Vary the sentence length on purpose. |
4. List what to leave out
Negative constraints are underused and they work. The model has strong default habits, and naming them directly is the fastest way to suppress them. This is the block I keep saved and paste into almost everything:
Prompt
| Avoid these words and phrases: leverage, robust, seamless, delve, unlock, elevate, game-changer, “in today’s landscape,” “it’s worth noting.” No em dashes. Do not open with a dictionary definition. Do not end by restating the intro. |
5. Ground the facts, or forbid them
Models will produce confident, specific, wrong numbers when you ask for data and give them none. Either paste the facts it is allowed to use, or tell it to write claims as general observations unless you have supplied a source. Both approaches beat hoping the invented stat happens to be right.
Prompt
| Use only the statistics in the notes below. If a point needs a number I have not given you, phrase it as a general observation instead of inventing a figure, and mark it with [check] so I can verify it before publishing. |
6. Draft in passes, never in one shot
This last one is a workflow rather than a single prompt, and it is where quality actually comes from. Approve the plan, then the draft, then run an edit pass. Steering three times beats rewriting once, and the sequence below is honestly ordered, so the numbering earns its place.
Outline first
Get the bones right before a single paragraph exists. It costs seconds and saves a full rewrite.
| Give me an outline only: working title, the one claim, and three section headers. No prose yet. |
Draft to the outline
Write in chunks so you can correct course between sections instead of at the end.
| Good. Now write section two only, following the voice sample I gave you. |
Run an edit pass
The draft is raw material. A tightening pass is what makes it read like a person wrote it.
| Now cut 15 percent. Remove any sentence that could open a different article. Tighten the weakest paragraph. |
The difference, side by side
Same topic, two prompts. This is roughly what the gap looks like once you put the patterns to work.
| The weak prompt | The engineered prompt | |
|---|---|---|
| What you type | “Write a blog post about remote team communication.” | 1,000 words for new managers leading their first remote team. Claim: most “communication problems” are really unclear ownership. Direct tone, voice sample attached, three H2s and one checklist, standing exclusion list applied. |
| Opening | A generic definition of remote work | A specific scene a new manager recognizes |
| Structure | Intro, three vague headers, summary | Hook, ownership framing, checklist, short takeaway |
| Facts | Plausible, unsourced statistics | Only supplied facts, the rest flagged |
| Editing needed | Rewrite from scratch | Trim and publish |
| The pattern underneath: the second prompt is not longer for the sake of it. Every extra line removes a decision the model would otherwise make badly on your behalf. |
Mistakes that quietly ruin the output
These are the habits that keep good prompters producing mediocre drafts. None of them are obvious while you are doing them.
• The mega-prompt. Cramming topic, voice, SEO, and structure into one dense paragraph. Split it into a plan pass and a draft pass so each instruction actually registers.
• “Make it SEO optimized.” That phrase gives the model nothing to act on. Hand it the actual target keyword and the search intent behind it instead.
• Skipping the voice sample. The single fastest fix for AI-sounding copy, and the one most people never bother to do. Two paragraphs is enough.
• Trusting the first draft. The first output is a starting point, not a finished piece. Publishing it untouched is how generated text ends up on the internet.
• Forgetting the exclusions. If you never say what to cut, the model keeps every habit you dislike. The list of things to avoid does as much work as the brief.
• Over-scripting it. The opposite failure. A rigid, line-by-line checklist produces stiff, joyless copy. Give clear direction, then leave the model some room.
A template you can copy
Here is the skeleton I paste and fill in. There is nothing magic about it. It just makes sure I never skip the parts that matter, on a tired Tuesday or otherwise.
Template
You are [role and stance]. Write a [length] blog post for [specific reader], who [what they know and want]. The single claim: [your angle]. After reading, they should [intent]. Structure: [hook / sections / table / takeaway]. Match the voice in this sample, and do not formalize it: [paste]. Use only these facts: [notes]. Flag anything unsupported. Avoid: [your word list]. No em dashes. No definition opener. |
Save it once. After a week of filling it in, the questions become second nature and you stop needing the scaffold at all.
The verdictAfter a couple hundred posts drafted this way, the lesson that stuck is dull and reliable: the model is only as good as the constraints you hand it, and writing those constraints is the actual work. The prompt is where you decide who this is for, what you are claiming, and what it must not sound like. Do that thinking up front and the draft comes back close. Skip it and you spend the time you saved on rewriting, plus the extra effort of stripping out the tells. I still edit every piece by hand before it goes out. Nothing here removes that step, and I would be wary of anyone who tells you it does. What these patterns buy you is a strong second draft in a few minutes instead of a weak first one in an hour. For anyone publishing on a schedule, that trade is close to the whole point. Start with the reader and the voice sample. Add the exclusion list. Draft in passes. The rest is refinement, and it gets faster every week you keep at it. |
