Handoff

Claude Code Handoff

Write one prompt the user can paste into a fresh Claude Code session. This skill is a transport fallback: it never launches Claude and never supervises execution. Return only the prompt. If the user is designing or reviewing the handoff itself, a short note before it is fine.

The recipient cannot see this conversation. Everything it needs in order to think from the same reality has to be in the prompt, and everything else is noise.

Claude leads the mission and chooses its verification. Codex is an optional /codex:rescue capability for a bounded need, not a default supervisor or decision-maker.

Ground it first

Look before writing. For coding work that usually means:

git status --short --branch
git diff --name-status
git diff -- <relevant paths>

Then read the files, tests, specs, ADRs, or logs the prompt will point at, and note which commands have already been run and whether they passed. For non-coding work, gather the equivalent source material. Skip the grounding pass only for something obviously small.

Write it

Carry the mission and the artifact wanted, the live state, the exact paths worth opening first, what you currently believe and why, and what is still open for the recipient to decide or re-check.

Two things are easy to leave out and expensive to lose: the evolution that explains the current direction (what the user reacted to, what was tried and dropped), and what the user will recognize as right beyond a passing test run.

Name real hazards: dirty user work, destructive git, deploys, migrations, security, licensing boundaries, dead paths to avoid, explicit non-goals. Leave out hazards that are merely conceivable.

Let length follow the mission. A small handoff is 12 to 20 dense lines; an ambitious one runs as long as it needs to. Cut transcript bulk, not the reasoning that got you here.

Say that Codex is available through the literal /codex:rescue command, since "Codex" alone does not name the command. For substantial work, suggest a couple of concrete seams (a diff to excavate, a wave to implement, a verification to run), and say plainly that they are examples.

The posture that makes it work

Prefer current read over settled decision. Prefer watch-out over prohibition. Prefer candidate seam over commandment.

The recipient is picking the work up, not executing a script. It should be free to reread the code, disagree with your conclusions, throw out the plan, and choose its own delegation shape. Say what you think is true and why, and do not promote that to settled unless an ADR, a code path, or an explicit user decision made it settled.

Close with the likely verification commands or evidence targets, and with what counts as done: a review memo, a PR-ready diff, a clean implementation branch, a verified command set, or a blocker list with the smallest remaining decisions.

For a one-line /goal, use agent-goal. A progress summary for the user is an ordinary response grounded in branch and session state, not a handoff artifact.