Your AI Setup Needs Spring Cleaning Too
A few months into using AI every day, you ask it for something simple and it answers with something you retired back in the spring, like an old service name or a rule you changed in June. You correct it, and a week later the same old detail shows up again from somewhere else.
THE BURROW.
I use the same picture with clients all the time. Think about the Weasley house in Harry Potter, the one that goes up and sideways because the family kept adding a room every time they needed one. Nobody went back to check whether the first floor could still hold everything above it. In a database, a tech stack, or an AI setup, that means nobody can find anything and every new addition leans on something that may not be holding weight anymore.
AI tools grow this way faster than anything I have worked with, because adding is so easy. A new skill takes ten minutes and a new project takes one. You never have to decide what the new piece replaces, so the old version stays in place and keeps answering questions.
Put the cords on the Counter.
When a client’s systems are tangled, the first thing I do is dump everything out. I put all the cords on the counter, we look at them together, and then we decide what gets plugged back in. That is where the Editor-Architect Framework starts: you edit the friction before you architect the flow.
So I did it to myself. I asked Claude to audit my skills, projects, memory, and scheduled tasks without changing anything. I wanted to know what overlapped, what was out of date, what contradicted itself, and what I kept doing by hand that should have a home. Then I asked for a ranked cleanup plan.
What I Found.
I tell clients all the time that no one knows their systems are messy until someone tells them. I was no exception.
Two of my skills were sending work to a third skill that did not exist. The word “decision” meant three different things across my daily and weekly routines, and two of those skills disagreed about where decisions should be logged. Twelve old projects still used my previous business name and a CRM I retired months ago, and any of them could show up in live work. My saved memory had two versions of the same fact in several places, written on different days by different conversations.
Each of those started as a small add-on that made sense when I made it. Together they explained a lot of the moments this year when I thought, this should not be this hard.
Read the report before you act on it.
When I went through the findings closely, some of what the audit called conflicts were choices I had made on purpose. I had removed due dates from certain tasks deliberately, and it flagged that as an error. It also flagged two rules that can both be true, like a daily cap on client work and a longer total workday.
I think of AI as an overeager intern. It will find everything and want to fix all of it, so you have to read what it hands you before anything changes. I sent the report back with corrections telling it which items were my decisions, which rules could both stand, which project to leave alone, and to mark anything it had guessed at. I acted on the second version.
Take it away before you put it back.
Most of the cleanup is taking things out. I am archiving the projects that belong to a business I no longer run and giving every kind of request one home. My scheduled tasks will call the skill directly instead of repeating it, so the two stop drifting apart. Where two facts disagree, I am keeping one. After all of that, I will build the few new pieces the audit showed were missing.
I also keep a list in ClickUp called Ideas, and once a month I let myself look at it to see which ideas still hold up. The same rhythm works for an AI setup. An afternoon every quarter is enough to keep it in order.
Erin Peshoff is the Founder of Vivid Operational Advisors. She has spent thirty years inside nonprofit operations, helped raise over $100 million for institutional missions, and built Vivid around the operating discipline most strategic engagements skip.
