# How do I hand work from ChatGPT or Claude to Codex without losing context?

_Hand work between AI agents with a durable brief containing the outcome, decisions, constraints, current artifacts and exact next action._

By Saad Ahmed, 2026-10-20

To hand work from ChatGPT or Claude to Codex without losing context, move the project state into a durable brief that both agents can read. Include the intended outcome, accepted decisions, constraints, current artifacts, unresolved questions and the exact next action.

Do not make the receiving agent infer the project from a pasted transcript. Transcripts contain alternatives and abandoned ideas without clearly identifying which choices won.

## Create a handoff packet

A useful handoff contains:

1. The outcome in one or two sentences.
2. What is in scope and explicitly out of scope.
3. Decisions already made and why they matter.
4. Links to files, branches, documents or designs.
5. Work completed and checks already run.
6. Open decisions that still require a person.
7. The next bounded action and its definition of done.

Keep source material linked rather than copied when possible. That gives Codex the current artifact and reduces stale duplicates.

## Store it with the work

For repository work, use a tracked project file or issue. Ask the outgoing agent to update it at the end of its session and the receiving agent to summarize it before making changes.

If the handoff crosses products and repositories, use a shared task board. The task becomes the neutral memory layer; ChatGPT or Claude can shape the brief, while Codex can execute and return evidence to the same record.

## Where ecco fits

ecco is agent-agnostic: connected assistants read and update the same tasks through MCP. The handoff stays attached to the outcome instead of belonging to one model's conversation history. Ownership can move from one agent to another while decisions and proof remain visible to the person supervising the work.

[Connect ChatGPT, Claude or another MCP client](/blog/connect-an-agent). To compare a shared board with keeping the workflow inside one assistant, see [ecco versus Claude and ChatGPT](/blog/ecco-vs-claude-and-chatgpt).

## Start Codex with a bounded instruction

```text
Read the task, project instructions and linked artifacts.
Summarize the outcome, settled decisions, constraints and next action.
Flag contradictions before changing anything.
Then execute only the approved next action and return evidence to the task.
```

The summary is a useful checksum. If Codex misunderstands the handoff, correct it before implementation spreads the error.

## Frequently asked questions

### Should I paste the entire ChatGPT or Claude conversation into Codex?

Usually no. Extract the final brief, decisions, artifacts and unresolved questions. Link the transcript only when its detailed reasoning is genuinely needed.

### What context is most often lost in an agent handoff?

The accepted option, rejected alternatives, non-obvious constraints, current artifact version and the distinction between completed work and suggested work are common losses.

### Can different agents use the same task board?

Yes, when each client can connect to the board and follows the same ownership and update rules. The board should remain the source of truth rather than any single chat.

### How do I check that the handoff worked?

Ask the receiving agent to restate the outcome, decisions, constraints and next action before it starts. Compare that summary with the durable task record.