Have the Previous AI Write the Hand-off Note── The receiver has not seen the conversation, so let the AI that holds it list the decisions and remaining work, and read the note before passing it on

01The operator did not write the hand-off; the previous AI did

In March 2026 the operator decided to move a piece of work to an AI running in a different environment from the one they had been talking to.

They did not write up the task themselves. Instead, they asked the AI they had been working with to put everything the other AI would need into a single note.

The operator pasted that note into the receiving AI, with one short line asking it to carry out the work described. The task continued from the note.

Two questions follow. What has to reach the receiving AI for the work to continue? And why should the previous AI write it, rather than the person?

02The receiving AI knows only the contents of the note

An AI in another environment has not read the earlier conversation. It does not know what was decided, or which approaches were tried and dropped.

In a 2025 engineering article on long-running agents, Anthropic notes that each new session begins with no memory of what came before. The same article describes leaving a progress file and a change history, which the next session reads first.

So the hand-off note becomes the whole of the receiver's starting knowledge. Anything left out of it does not exist on the receiving side.

03A receiving AI fills missing assumptions with guesses

When information is missing, a receiving AI usually does not stop. It supplies a plausible assumption from what it has and keeps going.

If that assumption differs from what was decided earlier, the mismatch shows up only after the work is finished. Fixing it starts with finding where the two diverged.

Anthropic's best-practice guide for Claude Code recommends a related sequence: have Claude write the specification to a file such as SPEC.md, then start a fresh session to build from it. The stated reason is that the build begins with a clean context and a written definition of done.

Medicine has long treated gaps at shift change as a safety problem. The patient safety primer from the US Agency for Healthcare Research and Quality defines a handoff as one provider updating another on a patient's status in order to transfer responsibility for care.

What moves in a hand-off is not only information. Responsibility moves too. That is why it matters to leave a written record of what was handed over.

04A hand-off note lets the receiver start and finish without the old conversation

Here is the definition used in this piece. A hand-off note is a document that lets the receiver begin the work, and judge for itself when it is done, without ever reading the earlier conversation.

The note has a boundary. It is not a summary of the conversation. A summary in the order things were said puts abandoned ideas and final decisions side by side with equal weight. The note records what stands at the end.

A task that fits in one sentence does not need a note. Restating the request is quicker. The note pays off when much was decided earlier and a receiver who did not know it could choose a different approach.

05In material review, it helps when reviewers or tools change

Pharmaceutical material review often moves mid-task to another reviewer or another tool. In each case, separate what the AI writes from what a person still decides.

In each case the previous AI holds the full history of the discussion. Pulling the basis for each comment and the exchanges with the authoring team out of that history is faster for the AI than for a person.

Whether something marked as agreed really was agreed is still a human judgment. The note's content becomes part of the review record, so a person reads it and confirms they can stand behind each decision before passing it on.

06The previous AI catches decisions but drops premises it takes for granted

Having the previous AI write the note works because that AI has the whole conversation in front of it. It can list the decisions more completely than a person working from memory.

Anthropic's 2025 article on context engineering explains that when Claude Code summarizes a long conversation for a new context, it keeps architectural decisions, unresolved bugs and implementation details, and discards repeated tool output. The same article warns that cutting too hard can lose small but critical context.

There are two weak points. The first is that the writing AI leaves out assumptions it treats as obvious. The more often a premise came up in the conversation, the more likely it is to be dropped as not worth saying.

The second is that it can write undecided matters as if they were settled. An idea that was only floated midway can appear in the note as a decision.

So the job does not end when the note is written. Tell the writing AI that the receiver has not seen the conversation, and have a person read the note before it goes out.

07From tomorrow: name the receiver, read the note, and have it played back

There are five steps.

1. Name the receiver. Say that an AI in another environment will do the work without reading this conversation, and ask for everything it needs.
2. Ask for four headings. Goal and done condition, what was decided, remaining work in order, limits and stop points. Have anything undecided marked as such.
3. Read it yourself. Check that every line written as a decision really was decided, and add the premises that were too obvious to be written.
4. Have the receiver play it back. After pasting the note, do not let the work start at once. First ask the receiver to state the goal as it understands it and the first piece of work it will do. The I-PASS handoff format in medicine ends the same way, with the receiver summarizing and confirming the plan.
5. Ask for the same note at the end. When the receiving AI finishes, have it write its progress under the same four headings. The next hand-off can start from that note.

Each hand-off costs a few minutes to read the note and a few more to check the playback. That is small next to the time spent tracing why finished work went the wrong way.