Look, we all know what AI writing looks like at this point. You've seen it. The weird formality. The bullet points nobody asked for. Words like "delve" and "tapestry" showing up where no human would ever put them. Sentences that start with "It's important to note that" when it really isn't that important and nobody needed it noted.
Opus 4.6 is genuinely capable of writing like a real person. The problem is that most people prompt it the same way they'd prompt any other model, and they get back the same sanitized, overly structured slop they've been getting for years. So here's what actually works.
Tell It What Not to Do
This sounds obvious, but it's the single most effective thing you can do. Opus 4.6 responds incredibly well to negative instructions. You don't just tell it to "write naturally." You tell it exactly which habits to drop.
Here's a prompt fragment that works surprisingly well:
That alone will strip out about 60% of the AI smell. Opus 4.6 is good at following these constraints without the writing feeling forced or stiff as a result. Lesser models sometimes overcorrect and produce something that reads like it's trying too hard to be casual. Opus mostly avoids that trap.
Give It a Voice to Imitate
"Write in a conversational tone" is vague and unhelpful. Every model interprets that differently, and the result is usually just the same AI writing with contractions sprinkled in.
What works better is telling it whose voice to approximate. You don't need to pick a famous author. You can describe a type of person.
Or for nonfiction:
The more specific you get about the persona, the less Opus falls back on its default patterns. Default patterns are where AI writing lives.
Break the Structure Addiction
Left to its own devices, Opus 4.6 will organize everything into neat sections with headers, subheaders, and numbered lists. That's fine for documentation but death for anything that's supposed to feel like writing.
Tell it directly:
You can also just tell it to write messier. That sounds counterintuitive, but real writing is a little messy. People go on tangents. They circle back. They repeat a key phrase for emphasis instead of restructuring to avoid repetition. AI writing never does any of that because it's optimized for clarity over authenticity.
The "First Draft" Trick
One of the best prompts I've found for natural-sounding output is telling Opus to write a first draft and not to clean it up.
Opus 4.6 is smart enough to understand that "imperfections" doesn't mean errors. It means human texture. It'll leave in a slightly redundant phrase here or a casual aside there. The result reads like something someone actually wrote rather than something that was generated.
Address the Hedging Problem
AI models hedge constantly. Everything is "can be" and "may" and "it depends" and "there are many factors to consider." Real writers have opinions and state them.
This alone makes an enormous difference. Opus 4.6 actually has pretty strong analytical capabilities, and when you give it permission to be direct, the writing gets dramatically better. It stops reading like a diplomatic press release and starts reading like something a knowledgeable person would actually say.
Kill the Intro-Body-Conclusion Format
Every AI model defaults to the five paragraph essay structure that was drilled into us in school. Introduction that tells you what you're about to read. Body paragraphs that make their points one at a time. Conclusion that restates everything you just read.
Nobody writes articles like that. Tell Opus to skip it.
This is especially important for shorter pieces. Nothing screams "AI wrote this" louder than a 400-word article that spends 100 words introducing itself and another 100 words summarizing itself.
Use Examples of What You Actually Want
If you have a piece of writing whose style you like, paste it in and tell Opus to match the voice. Not the content, the voice. The sentence length patterns, the level of formality, the rhythm.
Opus 4.6 is remarkably good at this. It picks up on subtle patterns that other models miss entirely. It can match the way a writer uses paragraph breaks for emphasis, or the way they drop in a one-sentence paragraph after a longer one to let a point land.
The Nuclear Option
If nothing else works and you're still getting AI-flavored output, there's one prompt addition that functions as a kind of reset button:
Is this a little heavy-handed? Yes. Does it work? Also yes. Opus 4.6 knows what AI detection tools look for and it can actively avoid those patterns when asked to. The output from this kind of prompt often feels more natural than most of the other approaches combined.
Putting It All Together
You don't need to use every one of these techniques at once. Pick the two or three that matter most for what you're writing and include them in your system prompt or at the top of your message.
The real takeaway here is that Opus 4.6 can write like a person. It's arguably the best model available for producing genuinely natural prose. But it won't do it by default, because its defaults are optimized for safety and clarity, not for sounding human. You have to actively steer it away from its trained habits, and when you do, the results are pretty remarkable.
The best prompt for human-sounding AI writing is ultimately a specific one. The more precisely you describe what you don't want and what you do want, the less room the model has to fall back on its defaults. And the defaults are where all the AI smell lives.
No comments yet. Be the first to comment!