Yes. You teach an AI agent from your edits by writing each correction down as a rule, storing those rules in one file the agent reads before every task, and reviewing that file once a week. The edit fixes today's output. The log fixes every output after it.
I learned that the slow way. For a few months my agents were busy, not useful. I would fix the same invoice format on Monday and watch it break again on Thursday, and each time I told myself the model was having a bad day. It was not having a bad day. It had no idea what happened on Monday.
Why does my AI agent keep making the same mistake?
Because nothing about your correction survives the chat window. Every new session starts from zero, so the agent meets you as a stranger, and you spend the first ten minutes explaining your business again. You are not prompting badly. You are running an agent with no memory layer.
This is the part most people skip. They treat AI like a search box: ask, get an answer, close the tab. But an agent that touches real work needs the same thing a new hire needs, which is context that persists after the conversation ends. That is the whole job of context engineering for founders. The phrasing of your prompt matters far less than the file your prompt reads from.
If you have ever said "every chat starts from zero," that is the symptom. The fix is not a better prompt. It is a written record.
What is an AI agent correction log, and what goes in it?
A correction log is a plain text file, usually markdown, that lives next to your agent's instructions and holds every fix you have ever made to its output. Mine is called corrections.md. It sits in the same folder as my project instructions, and it is the single most valuable document in my AI agent operating system.
Each entry gets three lines:
- What happened. One sentence, using the actual bad output. "Agent opened the vendor email with 'I hope this email finds you well.'"
- The rule. Written as an instruction, not a complaint. "Open with the reason for writing in the first sentence. No pleasantries."
- The check. How the agent verifies compliance before it hands anything over. "Before sending, confirm the first sentence names the specific invoice or deliverable."
A few real entries from my own file, so you can see the texture:
- Never write that I have been teaching for 30 years. Teaching since 1998. Business building since 1995. These are separate numbers and they never merge.
- Every invoice line item needs a project code.
- No exclamation points in business copy, ever.
For tools, the mechanics change but the pattern does not. Claude Code reads a CLAUDE.md file automatically. Cursor reads rules from .cursor/rules. A ChatGPT Project reads its custom instructions on every chat in that project. If you are not technical, a Notion page pasted into your agent's instructions at the start of a session does the same job.
Keep it in version control if you can. You get a diff every time you change a rule, which means you can see what you taught it and when.
How do you turn one edit into a permanent rule?
Stop editing and start classifying. Almost every correction I make falls into one of three buckets: tone, facts, or format. Tone is the hardest to write as a rule, so it needs an example of the bad version and the good version side by side. Facts and format convert cleanly into instructions.
Then do four things with the edit:
- Pair the negative with the positive. "Do not summarize the meeting. List the three decisions and who owns each one." Agents follow a paired example far better than a bare ban.
- Add the reason in under ten words. "Because clients read the subject line first." Models use the why, and so do you six months later when you have forgotten why the rule exists.
- Attach a check. A rule without a verification step is a suggestion. Give the agent one concrete thing to look at before it delivers.
- Date the entry and expire it. Rules rot. Every quarter I delete the ones that no longer reflect how the business runs, and I usually find three or four that are now actively wrong.
That last one is the part people miss. An agent that follows a stale rule is worse than an agent with no rules, because it is confidently wrong in a way you stopped checking.
How do you get the agent to actually read it before it starts?
Put the file where it loads automatically, and add one instruction at the top of every task: "Read corrections.md. Then list the three rules most relevant to this task before you begin."
That restatement costs about thirty seconds and cuts repeat mistakes more than any other change I have made. It forces retrieval instead of hoping the context window did the work.
Two more mechanics that matter. First, keep one shared file, not five copies across five agents. The moment you fork it, the copies drift and you are back to correcting the same thing in three places. Second, run a Friday review. Twenty minutes, scan the week's conversations, and promote any correction you made twice into a permanent rule. If you want a number to track for agentic workflow improvement, track repeat corrections per week. Mine went from dozens to a handful in about six weeks, and the drop was almost entirely from the review ritual, not from a model upgrade.
If you want the fuller architecture, including how the log sits next to your task files and handoff notes, it is in the AI OS playbook.
What changes once the agent stops repeating itself?
You get your attention back. I stopped dreading the invoice run and started handing off real work, and the reason was not that the AI got smarter. It was that the rules stopped living in my head.
The quieter benefit is the file itself. A correction log is an honest description of how your business actually runs, written by the person who runs it. When I onboard a human now, I start them on the same document. That is what teaching AI your business really means: not a clever prompt, but a written record of your standards that both your agents and your people read. That is why this pattern matters most for founders and operators, the people who are making every decision and cannot afford to make the same one twice.
FAQ
Can't I just use the memory feature in ChatGPT or Claude? You can, and it helps, but memory is a black box. You cannot read what it decided to keep, you cannot diff it, and you cannot hand it to a second agent. A correction log is explicit, editable, and yours. Use memory as a bonus on top.
How long should the correction log be? Long enough to cover your recurring work and short enough that you still read it. Mine runs to about four pages. If yours passes six, you are probably recording preferences instead of corrections, and the good rules are drowning.
Where should I store it if I am not technical? A Notion page or a Google Doc works fine. The requirement is that it loads into the agent's instructions at the start of a session and that you can edit it in seconds. Storage matters far less than retrieval.
I wrote the full playbook for this. How to Build Your Own AI Agent Operating System walks you through the exact architecture I use, step by step. You can get it at a.mastermindshq.business/ai-os-book.