Stuck on enterprise CoPilot but luckily I am on pilot ChatGPT at my day job. We are so locked down in CoPilot that I basically build my Knowledge base/ICM in Chat GPT, then create a single file markdown. The headers and sub headers with index in the file helps organize it enough for a single file CoPilot. So basically I have to drop certain sharepoint files into a copilot chat with some instructions. It is painful but it works.

@Jeffrey Taylor The thing I have been experimenting with is how much the AI used the headers/sub headers in Word docs. In markdown files it helps give the AI context of the layering of the information. #, ##, ### levels etc.

Markdown files are great for the headers and sections by nature. These give the AI out of the box some basic ideas on the hierarchical structure of the data. They are easy to store and diff on places like GIT.

@Johan Liebenberg What's worked for me is taking the code out of the check. On a client's booking system we split the testing. The client used the real app like a customer would (book, move, cancel), and a separate Claude session watched the database with read-only access and said out loud what each click actually wrote. It caught a sync hole the vendor docs never warned about: brand new appointments came through with no last-updated time, so the sync that looks for recent changes just never saw them. Reading the code wouldn't have flagged that. The data did. Two things make that independent, and neither one is the model. The checker starts fresh and gets the evidence, not the builder's summary, which is Bill's point. And it can read but can't write, same as Aurel's log, so the builder can't touch what the checker looks at. On different models, I'd change the brief before I'd change the model. A fresh session with no build history, told "find what breaks, don't fix anything," is your red team. My solo version of purple: the builder fixes, then one more fresh session walks the fixed version, because fixes break things too. On top of my own checks I've started running Cloudflare's free audit skill: https://github.com/cloudflare/security-audit-skill. It's built the way Bill describes, every finding goes to a fresh agent whose job is to disprove it. Their README says one run finds roughly half of what repeated runs find, though, and it only reads the code, so anything in your hosting or database settings comes back as a question for you to go check. Your mechanic analogy is where I end up too. AI is good at the pieces now. It'll find the bug, write the test, patch the hole. Where it's still weak is the shape of the whole thing: how it's designed and organized, and whether the parts are split up in a way that still holds when you add the next ten features. That takes someone who's watched systems break in production, had to fix them, and knows which patterns age well and which ones bite you later. For me that's the part still worth paying a human for. This is the lane I am leaning into since I have experience there.

Congrats Team 🙌

@Pascal Pollack I usually pull it all then sort the order out myself. Luckily my second job is not that busy yet :P

@Ruby Sparks I have seen some weeding robots