To correct an AI agent for good, turn any mistake you've seen twice into a clear rule it saves, with a reason and an example. If a rule doesn't help, the problem usually lies in the task, the missing context or the setup, and that's what needs changing.
Most people only ever edit the answer in front of them. That fixes today's draft, and next week the same mistake comes back, because an agent keeps doing whatever nobody has told it to stop.
Fixing one answer
The everyday correction: the reply is nearly right, so you change a sentence or two and use it. For a one-off, that's fine.
Before you move on, ask yourself whether you've seen this mistake before. If you have, editing again wastes your time, and it's worth setting a rule.
How do you make a correction stick?
Turn it into a rule the agent keeps, usually by saving it to memory. Instead of "make this shorter", say "Keep replies under 150 words unless I ask for detail." Instead of "that tone is wrong", say "I never use 'kindly' or 'please be advised' with clients. Plain phrasing only."
Rules hold up best when they follow the same advice model providers give in their own prompting guidance:
Be specific. "Be professional" gives the agent nothing to work with. "Short sentences, no jargon, no exclamation marks" does. Where you can, say what to do as well as what to avoid.
Give the reason. "Use bullets for summaries because I read them on my phone" hands the agent a principle it can apply to cases you didn't mention.
Show an example. One line of the right version teaches more than a paragraph of description.
Say when it applies. Some rules hold everywhere; others only for client-facing writing. A rule with no scope gets applied to everything, which causes problems of its own.
How should you word a correction?
Say what happened, what should have happened, and the principle behind it, in one calm message:
"That summary had six points. I need only the three that require a decision from me, because I read these on the way to meetings. If there are more, put the rest under a heading called 'for later'."
The agent can apply that to every summary that follows, without being told again.
What doesn't work:
- Sarcasm or frustration. It adds noise and no information.
- "Do better." There's nothing in it the agent can act on.
- Rewriting in silence. You end up with a good draft, and the agent has learned nothing.
- Five corrections in one message. Fix the most important thing first, then the next. Agents tend to apply long lists of corrections only in part.
When the setup is the problem
Some mistakes come from how the job was set up, and rewording the rule won't touch them.
Missing context. The agent can't know about a client it has never been told about. Add the facts, or expect generic output.
A task that's too broad. "Help with my admin" produces mush. "Draft replies to these three messages in this style" produces something usable. How to get the best use out of an agent covers narrowing a job down.
Conflicting instructions. You asked for brevity and also for a full analysis. The agent will pick one, possibly the wrong one, so settle it yourself.
The wrong model. A cheap model on a hard task fails in ways no amount of rewording fixes. Move the task to a stronger model; Choosing an AI model explains the trade-off.
An error inside a skill. If a saved procedure has a wrong step, every run inherits it. Open the skill and fix the step itself.
Why does an AI agent keep repeating mistakes?
Most often because it was never clearly told. After that, because the instruction was ambiguous, or because the correction was never saved.
Missing information explains far more repeat mistakes than anything resembling stubbornness. Ambiguity comes next: "short" means different things to different people, and to different models.
Then there's the correction that got buried. A rule mentioned once in the middle of a long conversation may not survive it, and an agent replying "noted" proves nothing. The Hermes Agent docs say this directly: unless the agent saves the note, "I've added that to my memory" is just text. State the rule clearly, then ask the agent what it remembers to check it was recorded.
Keep your rules in one place
Keep a short note of the rules you've set; ten lines is plenty. Rules pile up and some start to contradict each other, and a rule that made sense during a frantic month can hold the agent back later. Reading the list once a month catches both.
In a team, the same list turns one person's preferences into an agreed way of working. Teams that write their rules down and agree on them get far more consistent results from a shared agent.
When should you stop correcting?
When the mistake doesn't matter, when you've tried several fixes without success, or when correcting takes longer than doing the work yourself.
It happened once and doesn't matter. Let it go.
It keeps happening and nothing fixes it. The task is probably a poor fit for an agent. Some work needs a person, and forcing it means endless correction.
The output is fine, just not how you'd write it. Ask whether the difference matters. Writing slightly differently from you doesn't make the agent wrong.
Correcting takes longer than doing. Stop, and take it as a sign to rethink the task.
Even a well-corrected agent will sometimes misread a document, invent a detail or apply a rule too broadly. Corrections make that rarer without making it impossible, so anything important still gets read by a person before it leaves. Good rules shorten that review.
Common questions
Does an AI agent learn from corrections automatically?
Not always. A correction carries over only if the agent saves it, usually to memory or to one of its skills. Many agents do this when you state a clear rule, but a passing comment mid-conversation may be lost, and the underlying model doesn't change. Ask the agent what it remembers to confirm the rule was kept.
Should you start a new chat when an AI agent gets things wrong?
Sometimes. A very long conversation can make an agent lose track of earlier instructions, and a fresh start clears that clutter. A new chat won't fix a missing rule or a wrong memory, though. If the mistake returns in the new chat, correct the instruction or the stored note.
How do you stop an AI agent from making things up?
You can reduce it but not eliminate it. Give the agent the actual material to work from rather than asking it to recall facts, and tell it to say when information is missing. Ask it to point to where in your text an answer came from, and check names, numbers and dates yourself on anything that matters.
Can you undo a rule you gave your agent?
Yes. Tell the agent the rule no longer applies, ask it to remove the note from its memory, and then ask what it remembers to confirm it's gone. If the rule lives inside a skill, the skill needs editing too, or the old behaviour will come back the next time the skill runs.