← Back to all free guides
A free guide by Mariah Brunner ↗ INSTAGRAM ↗ TIKTOK

Stop Correcting AI
For The Same Thing

Four ways to make your feedback actually stick, starting with one prompt and ending with tools that save what you teach it.

"You’re absolutely right, I won’t do that again." Then it does it again. Here’s why: a promise in chat isn’t memory. Your correction only sticks if it gets written down somewhere the AI reads every time, as a short, specific rule.

Anthropic’s own Claude Code docs put it simply: "The more specific and concise your instructions, the more consistently Claude follows them."

Four Ways To Make It Stick

01

Any AI: Turn your correction into a clear rule

The moment it messes up, paste this. It writes your correction as an Always or Never rule with a good and bad example, follows it for the rest of the chat, and hands you a version to save.

Turn My Correction Into A Rule
You just made a mistake I don't want to see again. Turn my correction into a clear rule. The mistake: [paste what it did, or describe it] What I want instead: [describe it, or write "figure it out from my correction"] Do these 4 things: 1. Write the rule in one sentence. Start with "Always" or "Never." Say what to do instead, not only what to stop. 2. Show me one bad example (what you did) and one good example (what I want). Keep each one short. 3. Follow this rule for the rest of this chat. Before each reply, check your answer against it. If a reply breaks the rule, fix it before you send it. 4. Give me a copy-ready version I can paste into my project instructions. Use this format: RULE: [the one-sentence rule] WHY: [one short reason] DO: [good example] DON'T: [bad example] If my correction is unclear, ask me one question before you write the rule.

02

Any AI: Save the rule in your project instructions

Now it stops living in one chat. Every new chat in that project starts with your rules.

  1. Claude: open your Project, click Set project instructions, paste your rules and click Save instructions. Projects work on every plan, and Free gets up to five.
  2. ChatGPT: paste them into your Project’s instructions. They apply only inside that project and win over your general custom instructions.
  3. Keep the list short. Five sharp rules beat thirty vague ones.
Rules Block Template
## Rules (follow these in every chat) RULE: [Always/Never ...] WHY: [short reason] DO: [good example] DON'T: [bad example] RULE: [Always/Never ...] WHY: [short reason] DO: [good example] DON'T: [bad example] If a rule here conflicts with something I say in a chat, ask me which one to follow.

Want the full setup, with memory and global instructions too? See my 3-layer Claude memory system.

03

Claude Reflect catches your corrections

Claude Reflect notices when you correct Claude Code, like "no, use this instead," and adds it to a list. When you type /reflect, you see each one and pick Apply, Edit or Skip before it goes into your CLAUDE.md file. Nothing is saved without your OK.

Install it from your terminal (it needs Python 3.6 or newer), then close and reopen Claude Code:

Install Claude Reflect
claude plugin marketplace add bayramannakov/claude-reflect claude plugin install claude-reflect@claude-reflect-marketplace
  1. Work as usual. Start a message with remember: to flag a rule on purpose.
  2. Run /reflect --scan-history once to pull in corrections from past sessions.
  3. Good to know: Claude Code’s built-in auto memory already saves some corrections on its own. Reflect adds the review step and puts rules in CLAUDE.md, where you can see and edit them.

04

Claudeception saves what worked as a skill

Claudeception notices when Claude solves a hard problem the hard way, then writes what worked into a new skill. Next time the same problem shows up, Claude loads that skill and already knows the fix.

Install it with one command, then restart Claude Code:

Install Claudeception
git clone https://github.com/blader/Claudeception.git ~/.claude/skills/claudeception

After a tough fix, type /claudeception or say "save what we just learned as a skill." New skills land in your skills folder. Heads up: it’s built for coding work, it writes skills without asking first, and it hasn’t been updated since February 2026.

Not In Claude Code?

You can get the same "save what worked" habit in regular Claude. My Lock It In skill turns a workflow that finally worked into a reusable skill, no coding needed.

Using Claude Code a lot? This CLAUDE.md that improves from every mistake pairs well with Reflect.