> ## Content Index
> Fetch the complete content index at: https://www.jeffsu.org/llms.txt
> Use this file to discover other available public pages before exploring further.

# How to Clean Up Conflicting AI Instructions and Memories
- URL: https://www.jeffsu.org/newsletter-272/
- Published: 2026-11-20T15:05:00.000Z
- Updated: 2026-11-20T15:04:59.000Z
- Description: #272
- Author: Jeff
- Tags: #newsletter, ChatGPT, Claude, Gemini, #hide-tag

Want to see these tips early? [Subscribe to the newsletter!](https://www.jeffsu.org/newsletter/)

---

**Here’s a situation you’ve probably been in (at least once):**

1. Back in March, you got tired of long-winded answers from your AI, so you told your chatbot to “keep replies short,” and your AI “remembered” that preference.
2. Then in June, it skipped a step you actually needed, so you told it to “always explain your reasoning,” and the AI captured that feedback as well.

But here’s the problem: **Since BOTH of those instructions are still saved**, some answers now come out clipped, and others go on and on, and you’re like…wtf? **😑**

By the end of this post, you’ll **(1) know why this happens, and (2) have a 5-minute cleanup you can run today,** no matter which AI you use.

![Two instructions saved months apart, “Keep replies short” and “Always explain your reasoning,” each still shaping answers: one clipped, one that goes on and on.](https://storage.ghost.io/c/db/85/db858a9d-1c4f-4993-8042-faa87b9bb3c1/content/images/2026/11/AE272---Hook-Visual.png)

## Why your AI gets LESS consistent the MORE you correct it

Every time you correct your AI (e.g. “from now on, do X”), that correction gets saved somewhere you rarely (if ever) look: your custom instructions, its memory, a project’s instructions, etc.

**And nothing ever gets removed.**

So the more you use AI, the more hidden bloat piles up, the more “corrections” you try to make, and the vicious cycle continues.

![A loop: you correct your AI, the correction gets saved, nothing gets removed, instructions conflict, answers turn inconsistent, and you correct it again.](https://storage.ghost.io/c/db/85/db858a9d-1c4f-4993-8042-faa87b9bb3c1/content/images/2026/11/AE272---Vicious-Cycle-Visual.png)

Put simply: you never told your AI to check for conflicts, so **it’s perfectly happy keeping instructions and memories that contradict each other**, and that’s why you get inconsistent answers.

## The fix: delete or replace before you add

Before you give your AI a new instruction, ask the AI to take a look at what’s already there and work in this order:

1. **Delete** what’s outdated
2. **Replace** what conflicts or overlaps
3. **Tighten** what’s wordy
4. **Add** only when nothing existing can do the job

I call this “Defend Bloat,” and it’s the rule my own Defend Bloat skill follows (you can download it below).

![Typing /defend in ChatGPT brings up the Defend Bloat skill.](https://storage.ghost.io/c/db/85/db858a9d-1c4f-4993-8042-faa87b9bb3c1/content/images/2026/11/image.png)

💡

****Heads up:** the fix is DIFFERENT for those of you who are still using the web/chat version of ChatGPT, Claude, and Gemini vs. those of you who are using the desktop version. I’ve included tactics for both groups below. 😇

### If you use ChatGPT, Claude, or Gemini on the web

**Open a new chat and paste this prompt:**

```plaintext
Help me clean up the saved instructions and preferences that affect how you answer me. Start with the clearest conflict you can see. If you can’t see the original wording, ask me to paste the relevant entries from settings. Leave unrelated facts about me alone.

Show me the entries side by side and say where each comes from, if you know. If both could be useful in different situations, suggest wording that keeps both. Otherwise, ask which one I want to keep. Briefly flag clear duplicates or outdated instructions. If nothing needs changing, say so.

For each change I approve, give me the exact wording and where it belongs. Keep project rules in their project. Only give me a whole replacement for a settings field if you have its full current text and can preserve everything else. Change a saved memory only if I approve that specific edit. Tell me what you changed and what I need to do myself.

Include this rule once in my general instructions: “Before saving a new instruction or preference, check what you can see. If it repeats an existing one, keep the original. If they conflict, show me both and ask before changing either.”

Suggest one familiar task I can try with the old and new wording to check that the cleanup helped.
```

**Your AI can only flag what it can see, so treat what it finds as a starting point.** And once you’ve answered its questions, apply each approved edit where your AI tells you it belongs:

- **ChatGPT:** Settings > Personalization > Custom Instructions
- **Claude:** Settings > Your Claude > Instructions
- **Gemini:** Settings > Personal context

ChatGPT and Claude can remove or correct memories right there in the chat. In Gemini, you delete them yourself on that same Personal context screen.

Save the Defend Bloat rule once in your general instructions so your AI knows to check for conflicts next time.

### If you use the desktop app (and rely on .md instruction files)

If you use the desktop versions of Claude, ChatGPT, Gemini or anything else that reads a CLAUDE.md or AGENTS.md file, your AI can find the conflicts AND make the edits for you.

**IMPORTANT:** this skill needs an AI that can open and edit your files, so it won’t work on the web version.

[Download my defend-bloat skill](https://drive.google.com/file/d/1rjbxuY2jEDSXZu-8j9DtYfGxOAV92W4N/view?ref=jeffsu.org)

**Installation instructions (using Claude to illustrate):**

1. Left sidebar > Customize > Skills > Plus icon > Upload skill
2. Select the downloaded ZIP file directly (no need to unzip it).
3. Click Save, then check that defend-bloat shows as enabled in your Skills list.

It’s the same idea in ChatGPT Work/Codex and Google Gemini/Antigravity; heck you can even just attach the file and straight up ask the AI to explain the Skill in clear and simple terms and then install it for you (see how much more powerful the desktop versions are?).

![ChatGPT’s desktop app explaining that defend-bloat.zip holds one text file telling an AI how to clean up its standing instructions.](https://storage.ghost.io/c/db/85/db858a9d-1c4f-4993-8042-faa87b9bb3c1/content/images/2026/11/image-1.png)

**Once the skill is installed:**

1. Type “Use Defend Bloat on my CLAUDE.md” (or whatever your main instruction file is called).
2. It reads the whole file, shows you what it would delete, replace and tighten, proves nothing you still need gets lost, and waits for your approval before changing anything.

After that, type “defend bloat” at the end of a working session and it cleans up whatever changed.

💡

****Pro tip:** Google Antigravity recognizes BOTH `GEMINI.md` AND `AGENTS.md` so if you use ChatGPT Work and Antigravity for example, they can both read the same `AGENTS.md` file (reducing the risk of drift and bloat).

Next time you’re about to type “from now on,” ask your AI what the new instruction replaces first. Let me know how it goes! 😁

---

**Whenever you're ready, here are some other ways I can help you:**

🎛️ [The Cowork Academy:](https://coworkacademy.ai/?ref=jeffsu.org) Ready-to-install AI Command Center that loads your context automatically and makes every Claude session sharper than the last. 90% built when you download it.

💻 [The Workspace Academy:](https://academy.jeffsu.org/workspace-academy?ref=jeffsu.org) Never lose a file, task, or note again with my CORE workflow for Google Workspace. Adopted by over 10,000 Googlers.