Interrupting a Running AI Task: Put 4 Things in One Message── State that it is an interruption, the scope of the new request, the fate of the current work and the way back, and the two jobs are far less likely to mix

01The operator used a few opening words to mark a separate request

The March 2026 record holds a short request. With work in progress, the operator asked the AI to clear a different error. The message opened by saying this was an interruption, pasted about 20 lines of the error, and asked for it to be resolved.

That message settles two things. This is not a continuation of the current work. And the new job is to resolve the pasted error.

It leaves two things open. Should the current work stop or carry on? And once the error is gone, how does the original work resume?

The material for this episode holds only the request, not the AI's reply or the result. So we cannot say whether this message worked. What we can take from it is a form of writing, not a verdict. The opening words are worth keeping, and the two missing pieces are worth adding.

02The opening words tell the AI that a different job starts here

An AI answers with the conversation history in view. A new sentence with no warning may be read as part of the work already under way. Ask only for an error to be fixed, and the AI has to guess whether that error came from the current task or from somewhere else.

A short warning reduces that guessing. But it says only that this is a separate matter. It does not say how far the AI may reach or what becomes of the current work. So the message needs four parts, not one.

03Mixing the two raises the chance of error in the AI's output and in the person's return

The first problem is on the AI's side. A 2023 study reported that adding irrelevant information to grade-school arithmetic problems sharply lowered language model performance. The same paper lists an instruction to ignore irrelevant information as one way to reduce the damage. That study used word problems, not working conversations. It still gives a reason not to make an AI answer while carrying material that has nothing to do with the question.

The second problem is on the person's side. In one experiment, a sequence task was interrupted for an average of only 2.8 seconds, and errors in the order of steps doubled. At an average of 4.4 seconds, they tripled. People who are interrupted easily lose track of where they were.

Clinical work points the same way. A study observed 98 nurses giving 4,271 medications. Each additional interruption went with a 12.7% rise in clinical errors. The share of administrations with an error was 25.3% with no interruptions and 38.9% with three. This is an observed association, and it does not prove that the interruptions caused the errors.

Even when an AI does the work, a person checks the result and decides what comes next. If that person loses the place, a check gets skipped even when the AI's work is correct. That is the reason to write the way back into the message.

The third problem is cost. According to the Claude Code documentation, the whole conversation is sent with every request. A session carrying three unrelated tasks reads all three again on each turn of the fourth.

04An interruption is a request with a different goal, inserted midway

Not every message sent midway is an interruption. This episode defines it as follows. If finishing the new request brings the original work no closer to completion, it is an interruption.

One case is easy to confuse. Sometimes a separate error is what has stopped the original work. Then clearing it is a precondition for that work, and one line is enough for the way back: carry on once it is resolved. If the error has nothing to do with the original work, have a person check before anything resumes.

05In material review, it matters most when an urgent inquiry crosses products or materials

Here is a working scenario, not a recorded event. An AI is listing the revised passages in a material for product A when an urgent inquiry arrives about a material for product B. Ask in the same conversation, and the answer about product B is written in a conversation that still holds product A's indications, dosage and administration.

Someone must then check that none of product A's conditions slipped into the answer. When the product, the material or the version changes, a separate conversation is better than a warning in the same one.

Records give another reason to separate. In review, you must be able to explain later which comment applied to which version of which material. When requests mix, it becomes harder to tie an AI's output to that version.

A short question about the same product does not need a new conversation. Asking the name of a trial mentioned in the current document can stay where it is. The same rule works for literature summaries in medical affairs and for drafts in marketing.

The more urgent the request, the more it matters to read what you paste. An inquiry or an error message can contain patient data or unpublished information. Whether it may go into an AI is a question for your company's rules.

06Text sent during a task does not stop that task right away

Wording is only part of the difficulty. You also need to know how the tool behaves. As checked in the Claude Code documentation in November 2026, a message sent while the AI is working does not interrupt it. The message goes into a queue, and the AI receives it within the same turn, once the running operations finish.

So writing "stop" and sending it does not halt an operation already running. To stop, press the Esc key. The work stops there, and what has been done so far is kept.

These names belong to Claude Code, and other tools differ. Behavior like this can also change with updates. Try it once in your own tool and see whether a sent message stops the work. Opening a separate conversation is an option in most tools.

07From tomorrow: send one message with all four parts, and ask for a report before returning

There are four steps.

1. Decide whether it can wait. If it can, queue it. Interrupt only for what cannot wait.
2. Decide whether to separate. If the product, the material or the version changes, use another conversation. If the request stays in this conversation, go to the next step.
3. Stop, then record the state. If the current work must stop, use the stop control first. Then have the AI write a few lines on what is finished and what remains.
4. Send one message with all four parts. Include the interruption, the scope of the new request, what happens to the current work, and the way back.

Here is an example. "This is an interruption. Keep the current work stopped and leave its changes in place. As a separate matter, please resolve the pasted error. Do not touch anything you changed for the current work. When you finish, report in three lines what you changed and any effect on the current work, then wait for my reply."

The last sentence is the way back. If the AI returns to the original work on its own judgment, the person loses the chance to check the interrupting change. A report followed by a pause lets the person check that change and then tell the AI to resume.

When you resume, reread the state you had the AI record before the interruption. Then tell it to continue from there. If the interrupting change affected the original work, ask for that part to be reviewed as well.