Drift has causes. Name them.
- The conversation got long.
Early instructions get buried under later ones. The AI weights what is recent. Your rule from message three is now competing with two hundred messages of noise.
- The state lives in the chat.
If the only record of what was decided is the conversation, the AI has to remember it. It will not.
- Nobody checked.
The AI said it did something and everyone believed it. Unverified claims compound. By the fifth one the work is fiction.
- The problem changed.
You asked for one thing, then adjusted, then adjusted again. The AI is now solving the average of your requests.
The habits that keep it anchored.
- Fresh sessions for new jobs.
When the task changes, start a new conversation. Carry the state in a file, not in the chat.
- Write the state down.
A short file: what was decided, what is done, what is next. The AI reads it at the start. You update it at the end. This is the single biggest fix.
- Verify before you believe.
'Done' means shown, not said. Ask for the evidence: the output, the screenshot, the test result. If it cannot show it, it did not do it.
- Restate the job.
Every few passes, have the AI say back in two lines what it is building and what done looks like. If the restatement is off, you caught the drift early.
- Keep the rules in a file.
Standing rules live in the brain file or the skill, not in the chat. A rule in a file survives. A rule in a message fades.
Three signs it is already drifting.
- It starts using words you never used, especially confident ones.
- It stops asking questions it used to ask.
- Its summaries get more polished while the actual output gets thinner.
The summary read like everything shipped. The screenshot showed half of it. Ask for the screenshot first, the summary second.
Set up the two files.
Where the state lives, where the rules live, and what verification means for your work. Entries stay on this page and are not sent anywhere.
Re-anchor a session that has started to slide.
Paste this when the AI feels off. It forces a restatement and a check before any more work.
Stop. Before doing anything else: 1. Restate in two lines what you are building and what done looks like. 2. List every rule you are operating under. If a rule came from a file, name the file. If it came from this chat, say so. 3. List what you have claimed is done in this session, and for each one say whether you verified it (show how) or only believe it. 4. Say what you would change about the plan if you were starting fresh right now. Then wait. Do not continue until I confirm the restatement is right.
One prompt. Paste it into Claude. It builds the whole thing for your business.
The Pro pack is a build prompt with blanks for your context: your product, your medium, your team. Fill them in, answer its questions, and it produces the complete system in your words. The PDF explains the method so you can judge the output. The five kinds of drift and how each shows up, the guardrail file, the weekly audit, a regression set you can run in ten minutes, reset rules, and three prompts.
- The build prompt: paste it into Claude, fill in the blanks, and it builds the whole system for your business
- The five drift types: voice, facts, scope, format, decisions, with the early warning for each
- The guardrail file template that sits next to the brain file
- The weekly ten-minute drift audit checklist
- A regression set: ten canonical tasks with expected shapes, so you can test any tool or model in minutes
- Reset rules: when to start a fresh session and how to carry state over
- Three prompts: drift detector, style diff, fact check against the brain file
Delivered as a zip: the build prompt as a text file, the PDF, and every prompt and template as plain text. One payment, yours to keep.