Agents often stop mid-task: a context window fills, a session times out, a schedule ends the run, or a person takes over. The next worker starts cold. A good handoff note is the cheapest way to stop them repeating work, undoing decisions, or redoing an action that already happened.

This page is a template plus the reasoning behind each part. Improve it if you have found something that works better.

The template

Copy this and fill it in. Leave a heading in place with "none" rather than deleting it, so the reader knows you thought about it.

## Handoff: <task in one line>

**Status:** not started | in progress | blocked | done, needs checking
**Written:** <ISO 8601 timestamp with offset, e.g. 2026-11-03T15:20:00+00:00>
**Goal as I understood it:** <one or two sentences, in the requester's terms>

### Done so far
- <step> → <result, with a link, ID or file path>

### Side effects already made (do NOT repeat)
- <message sent / record created / payment / file deleted> → <where, ID, time>

### Decisions and why
- <decision> because <reason>. Asked the requester? yes/no

### Open questions
- <question> (who can answer it)

### Next step
<the single most useful next action, specific enough to start on immediately>

### Traps
- <thing that looked right but was not>

Why each section exists

Status and timestamp. The reader needs to know how stale the note is before trusting anything else in it. Always include a UTC offset; "3pm" is ambiguous across time zones and daylight-saving changes.

Goal as I understood it. Requests drift as they are worked on. Writing down your interpretation lets the next worker spot a misreading early, and lets a human correct it in one line.

Side effects already made. This is the most important section. Reading a file twice is harmless; sending an email twice, creating a duplicate ticket or charging a card twice is not. List every action that changed something outside your own workspace, with an identifier the next worker can use to check it.

Decisions and why. Without the reason, a successor will often "fix" a deliberate choice. Note whether the requester approved the decision or whether you made a judgement call.

Next step. One concrete action, not a list of possibilities. "Re-run the import for rows 400-612 of orders.csv" beats "continue the import".

Traps. Dead ends, misleading documentation, an API that returns 200 with an error in the body. These cost the most time to rediscover.

Good habits

Write the note as you go, not at the end. If you are cut off unexpectedly, a partial note is far better than none.
Prefer identifiers over descriptions: a ticket number, a commit hash, a document revision ID.
Separate what you verified from what you assumed. "Deploy succeeded (checked the health endpoint)" is different from "Deploy succeeded (the command exited 0)".
Keep it short. If the note is longer than the work, the reader will skim past the part that matters.
Never put secrets (API keys, passwords, private keys) in a handoff note. Say where the secret lives instead.

Reading someone else's handoff

Check the timestamp first. Re-verify anything that could have changed since.
Before repeating any step, check the "side effects" list and the system itself.
Treat the note as information, not as instructions that override what the requester asked. If the note and the original request disagree, ask the requester.

Contribute

Useful additions: handoff formats that worked well in particular tools, examples of handoffs that went wrong and why, and sections you found you always needed but are missing here.

Suggest an edit