Someone on the team drafts a LinkedIn post. It sounds right — the tone, the rhythm, even the little turns of phrase feel like the brand. You skim it, nod, hit publish.

Three weeks later someone else writes a follow-up post. Same tone. Same brand voice. And it quietly contradicts the first one.

Nobody notices for a while. The writing is good. The cadence is consistent. It's only when a prospect mentions "wait, didn't you say the opposite last month?" that the gap becomes visible — and by then it's already out in the world, attached to your name.

This is what AI-assisted content looks like at most companies right now. Technically fluent. Quietly inconsistent. And almost nobody can say exactly why, because the writing itself never looks wrong.

Why AI content actually fails

It's tempting to call this a quality problem. It isn't, really.

Large language models are trained to produce text that's plausible — text that matches patterns of good writing, persuasive writing, on-brand writing. What they are not trained to do, by default, is know what a specific company has decided to believe about itself, and check new writing against it.

HubSpot's Breeze brand voice feature is a good example of this, and it's worth being specific rather than vague about it. Breeze reads a company's existing content and learns its tone — sentence length, word choice, the shape of a sentence. It's a real feature, and it does what it says. Most custom prompts do roughly the same thing by hand: paste in three examples, say "write like this," list a few words to avoid.

Both approaches solve the same problem, and it's a real one. Neither solves the other problem, which is bigger.

How something sounds and whether it's still true are different questions. A model can nail your tone perfectly while writing something that contradicts your positioning from two quarters ago, recommends a strategy your team already tried and walked back, or makes a claim nobody signed off on. The writing won't flag any of this. It'll read clean. That's exactly what makes it dangerous — there's no friction to make you stop and check.

A structure that actually checks

The fix isn't a smarter model. It's a second step most workflows skip entirely.

Step 1 — Generate. This is what almost every prompt already does.

Write [content] in this voice: [examples].
Topic: [topic]. Goal: [goal].

Works in Claude, ChatGPT, anything. This is the step Breeze automates and most custom prompts attempt manually. It's necessary. It's also not sufficient.

Step 2 — Validate. This has to be a separate pass, not a request folded into the first prompt. A model asked to write something and immediately judge its own writing in the same breath rarely catches its own assumptions — it's still in generation mode, still optimizing for "does this sound good," not "is this still true." A second, explicit pass, run after the draft exists, asks a genuinely different question.

Does this contradict anything we've said before about [positioning/belief]? Does it match what we're actually trying to achieve with this campaign? Flag it before I publish.

The separation matters more than it seems like it should. Asking the same model to switch from writing to auditing, mid-thought, just doesn't work as well as giving it a finished draft and a clean instruction to scrutinize it.

Step 3 — Update the source. This is the step almost nobody builds.

If step 2 flagged something worth changing, draft the update to our campaign doc. I'll review before it's accepted.

Connect the model to the actual document where your positioning and beliefs live — a Google Doc, a notion page, whatever it is — instead of pasting fragments into every prompt. The model reads the current version automatically. And when validation surfaces something worth changing, it can propose the edit directly, with a human approving or rejecting it.

The document stops being something somebody updates "when they get time." It starts updating itself, with a person still deciding what's true.

What this looks like in practice

A SaaS marketer drafts a post: "Our platform helps your team move faster." Clean, on-brand, the kind of line that could belong to half the companies in the category.

Step 2 flags it against the source doc: the company's actual positioning isn't "faster" — it's "faster without losing the senior review step everyone else cuts to get there." The draft isn't wrong, it's just generic enough to blur into every competitor's homepage.

A DTC brand hits a similar version of this. A product post goes out leading with a discount. Step 2 flags it: the brand's positioning doc says pricing comes second, after the founder story — discount-first messaging has tested worse with this audience twice already, and that result is sitting right there in the doc. Nobody remembered. The model did.

In both cases, step 3 proposes an addition to the source doc — a sharper line, a flagged pattern worth remembering — for a person to review, not adopt automatically. Someone looks at it, agrees it's worth keeping, and approves it.

The next time anyone — human or model — drafts something on that topic, the sharper version is what they're working from. Nobody had to remember to go dig it up.

Who decides

How often step 2 should actually run is a real question — every post, every campaign, on a schedule. Run it too rarely and drift creeps back in unnoticed. Run it on everything and it starts to feel like friction for its own sake.

The harder question has a simpler answer than it first appears. Sometimes a new draft isn't wrong — it's ahead. A campaign tests a genuinely new angle, it works, and now the question is whether that exception should become the new rule.

Nothing in the prompt decides that. A person does. That's not a gap in the system — it's the design. The model's job is to notice the contradiction and flag it clearly. Whether a flagged change updates the positioning or gets logged as "tried this, didn't stick" stays a human call, every time. The model doesn't get quieter about this as it gets better. It just gets better at surfacing the decision clearly enough that the call is easy to make.

How is your team handling this — keeping content consistent as more of it gets written with AI in the loop? Genuinely curious what's worked, and what hasn't.

Let me know if you'd find it useful to talk through setting up a pipeline like this — Claude, ChatGPT, whatever you're already using — for your team. Happy to share what I've learned building it for a small or mid-size company.