Blog Inside Sneeze It · Inside the AI layer · Part 2
Correct it once, and every one of our tools learns
When someone on our team fixes something an AI wrote, that fix becomes a lesson six Sneeze It tools read before they write again. Here is how the Brain decides what counts, and what never goes agency-wide without a human yes.
Every agency that uses AI has the same quiet problem. A strategist rewrites a headline because the client hates the word "transform." Next week a different tool writes a report for the same client and uses "transform" again. The correction was real, it was right, and it lived in one person's head and one tool's history.
We wanted the opposite. Correct something once, in any Sneeze It tool, and every tool that writes for that client should know. So on October 5 we built the Brain: the agency's shared memory of what our AI got wrong and what the fix was. It went from first commit to live in one day, nine commits and about 2,000 lines of code, and it is now in its first week of use, inside the agency only.
Source: Sneeze Brain code, checked October 10, 2026
A contract with two clauses
The Brain asks very little of the tools around it. Each tool does exactly two things. It tells the Brain when a person corrects its AI. And it asks the Brain for lessons right before its AI writes. That is the whole contract.
Six tools follow it today: Studio (ads, emails, landing pages, video), Coach (weekly performance reports), the Hub (our record of every client location), Audience (market strategy), our call center tools, and Ideate (brainstorming). Each one carries the same small file. If the Brain is switched off or down, the tool behaves exactly as it did before. It never throws an error and never holds work up. We wanted learning to be something the tools gain, not something they depend on.
- 1A person correctsRewrites a headline, cuts a line from a report, scores an AI idea low
- 2The tool reports itOne signal, tagged with the client and the kind of work
- 3The Brain weighs itA person's word counts 1, a watched change counts half
- 4A lesson formsShort, tied to one client, one tool or the whole agency
- 5Every tool reads itBefore its AI writes the next thing for that client
Lessons attach to a location or a brand from the Hub, which is our one record of every client. That is what lets a lesson learned in Studio about a location reach Coach for the same location. A correction made in Studio guides Coach, and the other way round.
What counts as a correction
The Brain understands seven kinds of signal: someone said, the client said, a note on the client, asked to change, edited by hand, a check flagged it, and what happened.
They are not equal. A person stating something plainly ("never mention price in this client's ads") is the clearest signal there is, and it counts as a full point. A change we only watched, like a person redoing an ad or editing a line by hand, counts as half. A client lesson goes into use at one point. In plain words: one clear statement, or the same quiet change twice.
Source: Sneeze Brain learning rule
We made that trade-off on purpose. Watching edits catches the things nobody bothers to say. But a single hand edit could be a typo fix or a mood, so it takes two before the Brain acts on it alone.
Corrections also do not only happen inside our tools. Strategists say "never do this for this client" in Slack. So the Hub listens in a chosen staff channel. When a message is about exactly one client, it writes a plain restatement and sends that to the Brain. It never replies and never reads direct messages or channels shared with clients, and money is filtered out.
Lessons are guidance, never permission. Every tool still enforces its own rules after the AI writes.
The ladder to agency-wide
A lesson about one client stays with that client. But some lessons are really about us. If three different clients have taught the same thing, the Brain has Claude propose it as a lesson for a whole tool or for the whole agency.
Proposing is all it does. The wider lesson waits for a person to say yes. It may not carry any client's name. And a lesson that someone has turned off is never proposed again.
There are hard limits on what any lesson can be. A lesson can never be an offer, a price, a promised result, a customer quote or a private fact. Lessons steer how the AI writes. They cannot loosen the rules each tool already enforces. So a lesson can stop Studio from using a word your brand dislikes, but it can never put a discount into your ad.
What the team sees
The Brain opens on "What needs attention": six kinds of counted card over the last 30 days. Lessons waiting for a yes. The kind of work the AI is corrected on most. Clients who keep pushing back. Lessons that are not holding. Lessons that held. Tools that have gone quiet. Every number on those cards is a count of real rows. None of them are written by AI.
Once a day it checks whether there is enough material for a weekly read, where Claude summarizes the week's corrections into items, each with a "Do" action. There is also an ask box. Ask it "what do we keep getting wrong in reports?" and it answers only from what it holds, with the evidence. Give it a direction ("from now on, never...") and it drafts the lesson, which is saved only when a person clicks to confirm. Staff can reach the same thing from Claude or ChatGPT through our AI connector.
Before the Brain
A correction lives in one tool's history and one person's memory. The next tool repeats the mistake. Nobody can see which kinds of work the AI gets wrong most.
With the Brain
A correction becomes a lesson for that client in every tool. Patterns across three clients are proposed for everyone, with a human yes. The team sees counted cards of where the AI is weakest.
What comes next is joining lessons to outcomes, so we can tell which advice actually worked. The Brain already records what happened; using that to rank lessons is the next piece we are building. The same habit runs through our brainstorming tool, where the AI suggests and your people score.
