When AI gets something wrong in a chat, the damage usually ends with that chat. When the mistake gets saved into memory, a project file or your instructions, it becomes context, and every future answer builds on it. This is a 10-minute Sunday audit that catches those errors early and keeps your AI in sync with your life as it changes.
How one wrong assumption spreads
- It gets saved. A misread detail lands in memory or a context file.
- Future answers treat it as fact. Nothing flags it as a guess anymore.
- You add instructions to correct it. The original error stays where it was.
- Now the context contradicts itself. When two instructions conflict, Anthropic's docs note that Claude may pick one arbitrarily.
The more layers pile on top, the harder the cleanup. Fix the source, not every answer.
Errors aren't the only problem. Priorities shift, projects end and circumstances change. Context that was accurate a month ago can be just as misleading as a hallucination.
Where your context lives
Check whichever of these you use:
| Where | What to review |
|---|---|
| Claude memory | Settings > Memory, under Topics |
| ChatGPT memory and instructions | Settings > Personalization |
| Project instructions and files | Each active project in Claude or ChatGPT |
| Claude Code or Codex | CLAUDE.md, AGENTS.md, rules files and auto memory (/memory in Claude Code) |
| Your own context documents | Profile, priorities, current projects, preferences |
The 10-minute Sunday audit
Run it in one chat with access to your context files, or paste the relevant context in.
Step 1: Gather what changed (1 minute)
You need two inputs:
- Your current active context: paste it, or point the AI to the files.
- This week's notes and decisions: raw notes are fine.
Include changes to priorities, projects, tools, preferences, and anything you stopped pursuing.
No notes this week? Let the AI draft them from recent chats instead, and review the draft before you use it:
Based on our conversations this week, list what seems to have changed in my priorities, projects, tools and preferences. Mark every item as "stated" if I said it directly or "inferred" if you're guessing.
Step 2: Compare (3 minutes)
Compare my active context with this week's notes. List only what should be added, changed or removed, and quote the note that supports each change. Don't fill in gaps or add anything I didn't say.
Step 3: Find the conflicts (2 minutes)
Look for anything outdated, duplicated, contradictory or unclear.
Find conflicts in my context. Show each one side by side, name both sources, and ask me which version is current. Don't resolve any conflict yourself.
Asking you, instead of letting it choose, stops the audit from fixing one assumption by making another.
Step 4: Decide what stays (2 minutes)
| Decision | Use it for |
|---|---|
| Keep | Still current |
| Update | True, but changed |
| Archive | Useful history, no longer active |
| Delete | Incorrect or unnecessary |
Sort each proposed change and conflict into Keep, Update, Archive or Delete. Explain only the ones you're unsure about, then wait for my approval.
Step 5: Apply and verify (2 minutes)
Make the smallest clean change:
- Update the original source instead of adding a correction somewhere else.
- Remove duplicates.
- Move history to an archive instead of deleting it.
- Leave unrelated context alone.
Apply only the changes I approved. Edit the original source, remove duplicates and move archived items to an archive section. Then give me a short changelog and list anything still unclear.
Then test the result:
Using only my updated context, answer three questions: What am I focused on now? What is no longer true? Where are you uncertain?
If an answer is wrong, the source is still wrong. Fix it there.
Run it as one prompt
Once the steps feel familiar, combine them:
Run my weekly context audit.
Inputs: my active context [paste or name files] and this week's notes [paste].
1. List only what should be added, changed or removed, quoting the note behind each change. Don't fill gaps.
2. Show conflicts side by side with both sources and ask me which is current.
3. Sort everything into Keep, Update, Archive or Delete. Explain only uncertain items.
4. Wait for my approval, then apply only approved changes to the original sources.
5. Give me a changelog, then answer: What am I focused on now? What is no longer true? Where are you uncertain?
Keep it going
- Weekly: run this audit to capture what changed.
- Monthly: do a deeper cleanup of the setup itself: tools, skills, rules and servers. See How I Keep My Claude Code Setup From Drifting.
- Make it repeatable: save the prompt as a skill, or schedule a Sunday task that drafts the audit for you to approve. Never let it apply changes on its own.
Most people keep adding context. Very few maintain it. The order that works: update, compare, decide, verify.
Go deeper: How I Use AI to Index My Brain · Stop Building Your AI System Inside One Tool · Follow-Up Prompts That Fix Weak AI Answers