Short answer: The most important habit is redirecting instead of walking away. When the agent gets something wrong, tell it exactly what to fix and run it again. In Anthropic's research on roughly 400,000 Claude Code sessions, this was the single largest behavioral difference between people who got results and people who did not: beginners abandoned troubled sessions 19 percent of the time, experienced users only 5 to 7 percent.
If you build one habit for working with an AI agent, make it this: when something breaks, tell the agent what went wrong and run it again, instead of closing the window.
Anthropic's study of roughly 400,000 Claude Code sessions found this was the single largest behavioral difference between people who got results and people who did not. In sessions that ran into trouble, beginners abandoned the work 19 percent of the time. Experienced users abandoned it only 5 to 7 percent. Same tool, same kinds of problems. The people who succeeded simply did not treat the first wrong answer as a verdict.
Why a wrong first answer is not failure
An AI agent's first attempt is a draft. A wrong answer is not a dead end, it is information about what the agent misunderstood. The experienced user reads that and feeds it back. The beginner reads it as proof the tool does not work.
That reaction is the expensive one. The fix is usually one sentence away.
How to redirect well
- Say what is wrong, specifically. Not "that's not right." Try "you grouped the cancelled jobs with the completed ones," or "you used the created date instead of the payment date."
- Point to the rule it missed. The agent cannot see the exception you carry in your head until you say it out loud.
- Run it again. Then check the next version the same way you checked the first.
This is where your expertise becomes leverage. The same research found that knowing the work, not knowing how to code, is what predicts success. You are the one who can tell that the report mixed two kinds of jobs together. The agent cannot know that without you, and redirecting is how that judgment gets into the work.
It pairs with one habit on the front end: be specific before the agent starts, so there is less to correct later. That is the discipline behind the Test, Refine method for building an agent you can trust, and it is also the cure for the other reason beginners fail.
Stay in the chair. The first answer is rarely the last one, and the people who know that are the ones getting results.
Frequently asked questions
What is the most important habit when using an AI agent?
Redirecting instead of walking away. When the agent gets something wrong, tell it exactly what to fix and run it again. Anthropic's research found this was the biggest behavioral difference between people who succeeded and people who quit.
How is that different from what beginners do?
Beginners abandoned troubled sessions 19 percent of the time; experienced users only 5 to 7 percent. The experts did not hit fewer problems, they just kept steering instead of giving up.
What should I say when the agent gets something wrong?
Be specific. Name the exact error and the rule it missed, such as "you used the created date instead of the payment date," then run it again. Vague feedback like "that's wrong" does not give the agent enough to correct.
Keep reading: why your expertise beats coding skill, why beginners fail with AI agents, and the Test, Refine method.