Halfway through a bigger project, my AI often loses the plan we agreed to at the start. A financial plan. A product launch. A website move. A build. I start repeating myself, and handing it documents I already handed it.

I hit that earlier the day of the Roundtable. I told it, “I don’t understand why you can’t remember this stuff.” Then I stumbled on the fix.

“Lock it” after you say yes

There is nothing to lock until you have agreed. Research and argue first. When you would say “yes, that’s the plan,” stop and lock it. Do not wait until it has already wandered.

This is what I said, cleaned up only enough to paste:

Before that sentence, I had already told it to document the plan we agreed to. Now, it has a document to ensure it wont forget, and a duty to read that document, adding, changing, or checking things off every round moving forward.

This FIXED the issue!

Two parts. One document. The spec is what we agreed. The checklist is the path from here to done. Not two chats.

What locked looks like

On the call I used a website leaving an old agency. Fake names. Same shape for a launch, a long deck, or a deal.

If someone later says “add a redesign,” that is a new item it automatically adds to the list. It does not lose track. I didn’t have to tell it to add the new line. Once the rule is in, it adds the thing that came up.

Every reply ends the same way

Moving forward, at the end of every single response, it tells you what was just completed, whether anything changed on that document, and what’s next in the list. That is the whole trick. The last prompt did it, so the next one does it. It stops forgetting because you stopped letting a reply end without the list.

I don’t even add it to the project files for a job that lives in one chat. It keeps the document in that chat, and it has been told, every prompt, to update it. If the work is going to run for a long time, or across chats, then move the file into the project. Otherwise leave it where the work is.

The project is done when the last box is checked. Not when the chat feels finished.

Same lock, harder rule

I used the same sentence on a customer job where some scanned files contain private health information, and that information can never touch the cloud. We agreed. I said lock it. Checklist created. Then I asked how we get the work done without the private part ever being read by the cloud model.

The path we picked: the cloud model moves the file without reading it, then writes a program that runs on the desktop. The program finds the private columns and file names and cordons them off. The cloud model never looks at that part. It writes the programs. The programs do the private work locally.

That is not a compliance certificate. A cloud chat is still the wrong place for those files. The lock is what kept the rule from getting dropped halfway through the build. If you want the longer version of keeping files off the cloud, it is here: how to use AI on files you wouldn’t share.

Related: ask it to build the tool when the checklist should become a program, and make it find what it can do when it claims it can’t remember a step it can.