How to hand off work between AI sessions and models

Write a short handoff file before you stop: the goal, what is done, what is next, decisions with reasons, and files touched. Start the new session by having it read that file. Include why you chose things, not only what changed. A shared node can hold this record so you do not copy it by hand.

What a good handoff includes

A new session knows nothing. It cannot see what you tried, so it will repeat dead ends unless the handoff says what failed. Keep the file short enough to read in one minute.

Write it for a stranger. If a teammate who had never seen the project could pick it up from this file alone, a new agent can too.

Name the file the same way every time and keep it in the project root, so that any agent you open there can find it.

A template you can copy

Save this as HANDOFF.md in the project, ask the old session to fill it in, then open the new session with one line telling it to read the file first.

# Handoff
Goal:
Done and verified:
Written, not tested:
Next step:
Decisions (and why):
Tried and failed:
Files that matter:
Do not touch:
Check with:

Handing off between different models

Moving from Claude to ChatGPT, or from Codex to Claude Code, works the same way, because a Markdown file is readable by any of them. Do not paste the whole chat. Long transcripts waste the new model's context and bury the decisions.

Ask the outgoing agent for the summary, then read it yourself before you hand it on. A summary can drop a constraint, and the next agent will never know it was there. If the outgoing session hit a usage limit, write the file first, since you may not be able to ask afterwards.

Mistakes that make handoffs fail

Four mistakes cause most bad handoffs. Writing only what changed and not why, so the next agent undoes a decision. Leaving out failed attempts, so it retries them. Handing over a transcript instead of a summary, so the important line is buried. And trusting the summary without reading it. A two-minute read by you is cheaper than an hour of rework by the next session.

Checkpoints beat last-minute summaries

The best handoff is the one you updated along the way. Ask the agent to refresh the file after each finished step. Then a crash, a limit or a model switch costs you almost nothing. Commit the file with the work, or keep it beside the project, so it stays in step with the code.

For long-running tools, Anthropic documents a similar pattern for its memory tool: a progress log and a feature checklist that each new session reads first and updates last.

Where a shared node fits

The weak point of a handoff file is you. Someone has to write it, copy it and remember to point the new session at it. A NodePress node moves the record out of your hands. Agents that join the same node share a thread, a memory graph and the files you choose, and every step is written to one readable thread. A new agent can read what the last one did without a pasted summary.

It is still worth writing down your decisions and reasons. The thread shows what happened. Only you know what you meant.

Do it with a NodePress node

  1. Name a node on nodepress.ai. It is free, with no sign-up for the first one.
  2. Paste the one line into the session you are working in.
  3. Work as normal. Steps are written to the node thread.
  4. When you switch tools, paste the same line into the new agent.
  5. Ask the new agent to read the thread, then continue from the next step.

For a small one-off task, a five-line HANDOFF.md is quicker than setting up anything. A node pays off when you switch agents often or run several at once.

Questions people ask

What should a handoff file include?

The goal, what is done and verified, the next step, decisions with reasons, what failed, the key files, and a command to check the work.

How do I continue Claude Code work in Codex?

Write a handoff file, open Codex in the same project folder, and tell it to read the file first. Both tools read plain Markdown and the same project files.

How do I summarize a coding session for the next agent?

Ask the current agent to fill in a fixed template rather than write free text. Then read it and add anything it missed, especially the dead ends.

How do I resume work after hitting a usage limit?

Keep a handoff file updated as you go, so you can open a different agent and continue. If you wait until the limit hits, you may not be able to ask for the summary.

How do I checkpoint an AI agent?

After each finished step, write the state to a file or shared record and commit it with the code. A fresh session reads it and resumes from the next step.

Related

Sources

NodePress gives your AI agents one shared place on your own computer. Free to start, no sign-up. Make your first node.