I Told Founders to Delete Their AI Setup. I Hadn't Deleted Mine.

I’ve written twice now about AI setups rotting silently: once about why they rot at all, once about the test for which parts of an instruction file are already dead. Both posts told the reader to go back and check their own setup every 6 months. I hadn’t checked mine since I built it.

What I found in my own content pipeline

My site runs on a 5-agent pipeline: a strategist finds the angle, a researcher pulls the facts, a writer drafts, an editor cuts, a publisher formats for the platform. I wrote about building it back in May. What I hadn’t done since is check whether the agents still agreed with each other.

They didn’t. The writer and editor agents were each carrying their own hardcoded snapshot of my voice rules, typed directly into the agent file instead of reading my actual voice reference. That snapshot was last touched on 18 July. My real voice file has had three genuine additions since then: a full register system for writing differently across LinkedIn, blog, and client email, a set of side-by-side examples of what I actually sound like versus the AI default, and, five days before I went looking, a new rule about proving a claim with a real example instead of an adjective. None of it had reached the agents writing my content. They were still enforcing a shorter, older banned-word list and had no idea a register system existed at all.

The bigger one

Then I found the one that actually mattered. Sitting inside the live system prompt my sales agent uses to draft follow-up emails to real inbound leads was a single sentence: the company repo is public.

It isn’t. It never has been while that sentence sat there. It was true once, as a note to whoever was editing the file, got read into a customer-facing AI’s context verbatim, and nobody went back to check whether it still held.

Fixed the same day, live 18 days later

The easy part first. The content agents now point at the real voice file and the real distribution playbook instead of a copy typed in months ago, the same cascade pattern I use everywhere else: a short fallback list for the rare case the file can’t be read, the real file as the source of truth otherwise. The false line in the sales prompt is gone, replaced with what’s actually true. I wrote both fixes that afternoon, but they didn’t reach the live site until 28 September, so the false line sat in the sales agent for another 18 days.

Nick's voice rules:
- Short, declarative sentences.
- No these words: leverage, robust, seamless...

Before writing a word, read docs/voice.md in full.
Find the register named in the brief and write to that dial.
Re-read the draft against ai_default_vs_me before returning it.

That part took an afternoon. It’s mechanical once you’ve found it, which is exactly why it sat there for two months. Nothing broke. The output stayed technically on-brief. It just quietly stopped being as good as the file it was supposed to be reading.

The harder question this opened

The AI Brain framework has been published in the open on this site for a while: the method is not the moat, so I put the whole thing on a page anyone can read. The source repo behind it, the actual prompts, the actual agent files, the actual commit history, is still private.

Those are two different claims, and I’d been treating them as the same one. Publishing the method is a page. Building in the open is a different, bigger decision: what a live audit like this one looks like in public, what a real commit history says about how the business actually runs week to week, and whether that’s a stronger proof of the work than another page describing it. I think it probably is. I’m not flipping the repo to public on that instinct alone, there’s a proper sweep to do first for anything client-specific sitting in the history, and it’s the kind of call worth pressure-testing rather than deciding at 11pm because it made a good closing line.

What I can say now: the agents writing this post are reading the real file. The sentence that wasn’t true is gone. If you’re running anything that drafts on your behalf, worth going back and checking whether it’s still reading the version of you that’s current, or the one you were three months ago. If you want that audit run properly on your own setup, that’s the first afternoon of an AI Brain engagement. If you’d rather talk it through first, book a call.