When you ask an AI to "expand" a blog draft, you're asking for two very different things at once. You want more material — more examples, more context, more polish. But you also want it to still sound like you, make your point, and not wander into territory you never intended. Most people discover the hard way that an AI is better at the first job than the second.
The failure is rarely the model's fault. It's a scope problem. If you hand the whole draft to the model and say "make this longer and better," you've given it permission to touch everything — including the parts that were already good, and the parts that carry your actual point of view. It will happily "improve" your central argument into something generic, because generic reads as "correct" to a model that has no stake in your opinion.
I use a simple way to prevent this. I call it the three-zone rule. Every paragraph in a draft gets assigned to one of three zones before I ever open the AI. Zone one is yours and yours alone. Zone two is collaborative — the AI drafts, you verify. Zone three is the AI's to handle with minimal oversight. The rule doesn't make AI smarter. It just makes sure the AI only works on the material where it actually helps.
Zone one: the parts only you can write
Zone one is the smallest zone, and the most important. It holds everything that carries your personal stake in the piece: your own stories, your opinions, your direct experience, your conclusions, and any claim that depends on your credibility rather than on public facts.
Why keep these out of the AI's hands? Because a model doesn't know what you actually believe. It knows what people in general believe about a topic. When it rewrites your opinion, it doesn't improve it — it averages it. The sharp, specific take that made the piece worth reading gets smoothed into a take everyone already agrees with. You lose the one thing an AI can't generate for you.
This is also where the "don't fabricate real people" rule lives naturally. Your zone-one stories involve real situations — a client you worked with, a talk you gave, a mistake you made. You know what actually happened. The AI doesn't, and if you let it "expand" a story, it will quietly invent details to make the story tidier. A story that didn't quite have a clean ending will suddenly gain one. A client who had no name will suddenly become a specific person with a specific business.
The practical rule for zone one is blunt: the AI does not rewrite this text, and it does not add new sentences inside your stories. You can ask it to check for typos. You can ask it to tell you if a sentence is confusing. You cannot ask it to "improve the flow" of a story, because flow-improvement is exactly where invented detail sneaks in.
Zone two: the AI drafts, you verify
Zone two is where the expansion actually happens, and it's the zone most writers underuse. This is the material you need more of, but where the facts are checkable and don't depend on your personal experience: definitions, background context, statistics, comparisons, lists of tools, and step-by-step explanations of how something works.
The key word is verify. AI is very good at drafting zone-two material quickly, and very capable of being confidently wrong about it. It will cite a statistic that sounds plausible and isn't real. It will describe a feature that doesn't exist. It will invent a year for a product launch. So zone two has a firm rule: nothing the AI writes here goes into the article until you've confirmed it against a source you actually saw.
This doesn't have to be slow. For most blog material, a quick check is enough. If the AI says a tool has a free tier, you spend thirty seconds on the tool's pricing page. If it quotes a number, you do one search and look at the result yourself. If it explains a concept, you read it back and ask yourself whether that matches what you know. The point isn't to make AI useless — it's to make AI a fast first draft of the boring parts, with you as the editor who signs off.
The reason zone two is collaborative rather than "AI-only" is trust. When a reader catches one wrong fact in your post, they stop trusting the rest of it — including your zone-one opinions, which were genuinely yours. A single invented statistic can discredit an honest argument. So the verification step isn't busywork. It's what keeps the AI's errors from contaminating the parts that matter.
Zone three: the connective tissue AI can own
Zone three is everything that holds the piece together but doesn't carry meaning on its own: transitions between sections, subheadings, bullet-point formatting, meta-descriptions, and the many small rephrasings that turn rough notes into readable prose. This is the zone where you should let the AI work freely, because getting it wrong here is cheap and easy to catch.
The reason zone three is safe is that it's low-stakes by definition. If an AI writes a slightly awkward transition sentence, you notice it immediately when you read the piece back, and fixing it takes five seconds. If an AI proposes a weak subheading, you can reject it and ask for another. None of this touches your argument or your facts, so there's no hidden cost to letting the model run.
In practice, zone three is also where you save the most time. Manually writing transitions and reformatting notes is tedious and mechanical. Delegating it to the AI frees up the mental energy you need for zone two verification and zone one writing — the parts that actually determine whether the piece is any good. The trick is simply to be explicit about what belongs in this zone, so the model doesn't treat zone three as an invitation to expand zone one.
Putting the rule into a single prompt
The whole framework collapses into one short instruction you can paste at the start of any expansion request. It goes something like this:
I'll paste a draft. Treat it in three zones. (1) Do not rewrite or add to my personal stories, opinions, or conclusions — you may only flag typos or confusing wording. (2) For background, definitions, statistics, and step-by-step explanations, you may draft new material, but mark every claim that needs verification so I can check it. (3) You may freely improve transitions, subheadings, and formatting. Expand only zones two and three.
That's it. When I paste this before a draft, the difference in output is immediate. The AI stops "improving" my argument and starts doing the two jobs it's actually good at: generating checkable background material and cleaning up connective tissue. My voice survives, and I still end up with a longer, better-supported piece.
What's happening under the hood is worth noticing. You're not using a cleverer model or a fancier prompt. You're just removing ambiguity about scope. Most "the AI took over my draft" complaints come down to the model making a reasonable but wrong guess about what "expand" meant. The three-zone rule replaces that guess with an explicit contract.
What "getting steered off course" actually looks like
It helps to know the specific failure modes, because they're easy to miss until you've shipped a post and read it back with fresh eyes. Here are the three I see most often, and how the three-zone rule catches each one.
The first is thesis drift. You start with a specific claim — say, "AI helps solo bloggers more than it helps big teams" — and the model expands it into a balanced, both-sides paragraph about how AI helps everyone. Balanced reads professional, but it's no longer your point. The three-zone rule prevents this because your thesis lives in zone one, and the model never touches it.
The second is invented specifics. The model, trying to make a story more vivid, adds a concrete detail you never mentioned: a number, a name, a location, a timestamp. This one is dangerous because invented specifics are what turn an honest anecdote into something that could embarrass you if it's wrong. Keeping stories in zone one — no rewriting, no additions — is the only reliable guard.
The third is style drift. The model starts writing in a voice that isn't yours — more corporate, more American, more "LinkedIn." This is subtler and doesn't look like an error. The fix is partly the three-zone rule and partly a separate instruction: tell the model what your voice sounds like, or better, give it a short example of your writing and ask it to match. I keep a two-paragraph voice sample saved for exactly this.
A real workflow, start to finish
Here's how this plays out for an actual post, so you can copy the sequence. Let's say I'm writing about using OCR to turn receipts into searchable notes — a topic I've actually worked through.
I write the rough draft myself, in a plain text file, without touching the AI. That draft has my story (the drawer of receipts I finally digitized), my opinion (OCR is worth it mainly for search, not for storage), and my conclusion. That's zone one, and it's maybe forty percent of the final length.
Then I paste the draft with my three-zone prompt. I ask specifically for zone-two additions: what OCR actually is, how the accuracy differs between printed text and handwriting, what free tools exist, and any workflow caveats. The model returns those sections, and it flags the claims I should check — tool names, accuracy claims, pricing.
I spend ten minutes verifying. One tool name turns out to be discontinued. One accuracy figure is too old. I fix or delete those, keeping the rest. Then I ask the model for zone-three work: tighter subheadings, better transitions, a bulleted version of the workflow steps, and a meta-description.
The final read-through is where I catch what slipped through. Almost always, the slips are in zone three — an awkward transition or a subheading I don't like — and they take seconds to fix. The zone-one story still reads like me, because the model never rewrote it. The zone-two facts are things I actually checked, so I'm not worried about being wrong in public.
Why the zones need to stay separate in your head
There's a temptation to blur the zones because the model is good enough to fool you. It can write a plausible version of your own story. It can draft a statistic that sounds right. It can smooth a transition so well you don't notice it rewrote your point.
The three-zone rule isn't really about the AI. It's about keeping the categories clear in your own mind before you paste anything. Once you've decided that a given paragraph is zone one, two, or three, the instructions to give the AI become obvious, and the verification steps become obvious. The discipline is upstream of the prompt — it's a decision you make while writing the draft, not while editing the output.
That's also why the rule sticks. It's not a technique to memorize or a long prompt to maintain. It's a habit of asking, for each paragraph, "who owns this — me, both of us, or the AI?" Answer that once, and the rest follows.
Where the rule breaks down
It's worth being honest about the limits, because no framework survives contact with a real deadline.
The three-zone rule assumes you have a real point of view and a real story to protect. If you're writing a purely factual, low-opinion post — a tool comparison, a how-to with no personal angle — zone one nearly vanishes, and the post becomes almost entirely zone two and three. That's fine. The rule still works; you just spend more time verifying and less time defending your voice.
The rule also adds a small overhead on every post: you have to tag paragraphs, write the scope prompt, and do the verification. For a short, time-sensitive post, that overhead can feel like more than the post is worth. My answer is to scale the rigor to the stakes. A quick news roundup gets a light version — one-line scope prompt, a fast fact-check. An opinion piece that carries my name and reputation gets the full treatment. The rule isn't a straitjacket; it's a dial.
The one place the rule genuinely fails is when you have no draft at all and ask the AI to write from a blank page. Then there's no zone one to protect, and the result is almost always generic, because the model has no stake in it. The three-zone rule can't fix a missing author. It only helps when there's already a you in the text.
Start with a smaller version
If you want to try this without overhauling your whole workflow, start with just the zone-one rule: before you ask an AI to expand anything, take sixty seconds to mark which sentences are your stories, opinions, and conclusions, and tell the model not to touch those. That single instruction fixes the most damaging failure mode — the silent flattening of your argument — and it costs almost nothing to adopt.
From there, adding zones two and three is just a matter of labeling the rest of your draft. And once you've seen your voice survive an expansion pass intact, you won't go back to the "make this better" free-for-all. You'll have felt the difference between an AI that helped you say more, and an AI that quietly said less on your behalf.
