You’ve been lied to. The AI industry wants you to believe that if your generated documents sound like a robot wrote them, it’s because your prompts aren’t good enough. That’s convenient — because it shifts the blame to you. But the truth is nastier: some AI models are fundamentally incapable of writing like a human. No amount of prompt engineering will fix an architecture that treats every sentence like a legal disclaimer.
I learned this the hard way. Last week, I fed Grok 4.6 a stack of project documents — the kind of messy, real-world specs that any product manager deals with. I wanted a clean summary. What I got was a 6,000-word monstrosity that repeated the same point three times, used phrases like ‘in the context of the aforementioned scenario,’ and somehow managed to make a simple process flow sound like a government regulation. It wasn’t wrong. It was just unreadable.
And that’s when it hit me: The biggest lie in AI-assisted writing is that more context equals better output. The opposite is true. The more you feed a model, the more it tries to cover every edge case, every possible exception, every historical footnote. It doesn’t know that your team already agreed on the basics. It doesn’t know that the document from three months ago is obsolete. So it strains to be ‘comprehensive’ — and ends up being exhausting.
This isn’t a prompt problem. It’s a model problem. Some models are built to reason, not to communicate. They’re translation-ese machines: they can handle logic puzzles but can’t tell the difference between a specification and a story. The solution isn’t to write better prompts. It’s to choose a model that actually speaks your language, and to starve it of the garbage it would otherwise treat as gospel.
I’ve since tested this across five models. Claude 4.6? Smooth, natural, almost human. GPT‑5 fluctuates — sometimes brilliant, sometimes hollow. Grok? A disaster for anything meant to be read by a human. DeepSeek? Fine for code, mediocre for documentation. The pattern is clear: You can’t make a bad writer into a good one, no matter how many instructions you give them. The model’s DNA decides how it handles language. Pick the right one for the task.
But even the best model will fail if you feed it everything. I’ve learned to ruthlessly prune the context. I now tell the AI exactly which files are historical, which are obsolete, and which are the current rules. I write an AGENTS.md file that says: ‘Ignore anything in the /archive folder. Only read the /current folder.’ That single change cut my output length by 40% and doubled readability.
Another trick: let the AI write the first draft, then re‑write it yourself. Use the model for structure and logic, not for final prose. AI can help you say it better, but it can’t decide what to say. That’s still your job. And if you hand that over, you’ll get a document that’s technically perfect and humanly dead.
So next time you’re staring at a bloated, robotic AI output, don’t tweak your prompt. Ask yourself: is this model built for people, or for accuracy? And did I give it a diet of noise it swallowed whole? Fix those two things, and you’ll get documents people actually read.
FAQ
Q: But if the model is the problem, why do some people get great results with the same model?
A: Because they’re doing two things differently: they prune the context so the model doesn’t get distracted by noise, and they choose the right model for the task. A model that excels at code generation may be terrible at writing prose. The same model can produce wildly different results depending on what you feed it and how you post‑process it.
Q: What’s the practical takeaway for a product manager who uses AI daily?
A: Stop treating AI as a writing assistant that just needs better instructions. Instead, treat it as a tool with specific strengths. Use a reasoning model to structure your thinking, then switch to a language‑focused model (like Claude) for the final output. And always, always restrict the context — the AI doesn’t need to see every old document to write a new one.
Q: Isn’t this just a case of ‘garbage in, garbage out’? You’re just blaming the tool instead of the user.
A: Partly true — but the ‘garbage in’ isn’t your prompt, it’s the model’s architecture. Some models are inherently verbose and can’t distinguish between a spec and a story. You can write the perfect prompt and still get robotic output if the model’s training data or architecture biases it toward over‑explanation. The industry has oversold prompt engineering; the real leverage is in model selection and information diet.