# What it takes to keep a marketing brain true > Source: > Published: 2026-11-05 10:21:11.417134+00:00 # What it takes to keep a marketing brain true In [a recent Growth Unhinged](https://www.growthunhinged.com/p/ai-brain-for-your-positioning), Melissa Kwan describes how she repositioned eWebinar with an AI brain. It’s a Claude project holding call transcripts, customer quotes, objections, competitor cheat sheets and the company’s old copy, and she used it to rewrite everything from the homepage to the demo script. Time on their site went up by half, and they closed their first enterprise deal. It’s the most concrete account of the method we’ve read, and we’d underline her conclusion that the thinking has to stay yours. We run our marketing on the same kind of brain, and have for months. Ours is a git repository of Markdown documents that Claude Code reads before it writes anything. They cover what the product does today, how we position it, a profile of every competitor, every acquisition channel we’ve tried and why we stopped, and the voice rules our copy has to follow. Building it was the easy part. Keeping it true is the hard part. Her piece touches on it where she puts every finished page back into the project and updates the prompt. That’s the right instinct, and it’s also where a brain starts to drift. Four rules came out of it for us. ## Write it down before you answer In August we pasted ten notes from a conversation with a sales leader into a session. We discussed the notes at length, but Claude didn’t store them in a file. Later when we needed them, they were nowhere to be found, so we pasted them in again. Acknowledging something is not the same as recording it, and it’s worse than saying nothing, because whoever supplied it reasonably assumes it’s stored and stops repeating it. So the rule became that Claude writes anything we tell it to a document before it replies, not after. The test is simple enough for anyone to apply: did a file change? ## A correction you give twice becomes a check A brain learns your voice from what you tell it, but it forgets just as reliably. We found ourselves making the same corrections to copy over and over. A “carry” metaphor here, a sentence that narrated what the page was about to do there, a claim framed by what something isn’t. Writing the correction into the instructions helped for a while. What fixed it was turning each repeated correction into a check that runs when the site builds. Each banned phrasing cites the correction that earned it a place, and a new or edited page that uses one fails the build. Now when we have to correct Claude the same way twice, the correction becomes a check. ## Decide which document wins Melissa’s project instructions tell the AI which source wins when two disagree, and that’s the line in her piece we’d point other founders to. We learned why it matters the hard way. A website audit we’d written said we weren’t listed on any review sites, while the document that tracks our results said we were listed on G2, Capterra and Trustpilot. The audit was simply wrong, but until we said which document to believe, the brain had no way to choose. Now the brain’s instructions name the authoritative document for each subject, and when another document contradicts it, the authoritative one wins and the other gets fixed. ## Read the whole thing, every time The subtlest failure came from editing one section at a time. Claude revised a comparison post section by section, and each section read fine on its own. But one clause said both products stopped at the mailbox, and the next said the other product covered mail, calendar and meetings. Nobody had reread the whole post, so nobody saw the two disagree. So a second check now reads each comparison as a whole and looks for exactly that: a “both” or a “neither” that a later sentence contradicts. And the person reviewing a draft reads it once, start to finish, before calling it done. The checks spare the person rote fact-checking, so they can focus on style, tone and the strength of the argument. ## Most of what goes in arrives in messages Look at the inputs Melissa lists, and most of them first arrived in someone’s inbox or chat: the customer quotes, the objections, the reasons people chose a substitute instead. Ours are the same. Last month a competitor announced a rebuild of its product in an email, and our comparison pages were wrong until someone noticed, pasted it in and asked for an update. That’s the gap in any brain kept current by hand. These rules make sure that what reaches the brain gets recorded, checked and believed in the right order. But the brain’s accuracy still relies on someone giving it context. That’s why analyzing messages to update the brain is so valuable. this+that reads a team’s messages as they arrive, and its workflows write what they learn to a shared brain, with a person approving each update where you want one. We’re building toward one workflow that keeps most of the brain current. We still run our marketing from the repository, but we’re planning to move all of it onto the brain in this+that, so the competitor emails, the customer quotes and the corrections reach it without anyone remembering to paste them. Melissa’s advice holds either way. Decide the direction yourself, give the brain everything, and keep it current.