Handoff

Handoff

Write one prompt the user can paste into a fresh session. Use the destination the user names; otherwise keep it agent-neutral. This skill prepares the handoff and does not launch or supervise the recipient. Return only the prompt. If the user is designing or reviewing the handoff itself, a short note before it is fine.

Give the next session enough context to make the next useful judgment. The recipient cannot see this conversation: preserve the intent, evidence, and reasoning it would otherwise lose, and point to accessible sources for details it can recover. Do not assume tools or delegation arrangements from the recipient's name.

Ground it first

Look before writing. For coding work that usually means:

bashCopy code
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 distinguish work completed from checks actually run, their results, and what remains unverified. For non-coding work, gather the equivalent source material. Skip the grounding pass only for something obviously small.

Match the prompt to the recipient's known access. Identify the checkout and relevant paths when it can read them; otherwise include the excerpts or evidence needed for the next judgment. When access is unknown, include the essential context and name any missing source or tool needed to proceed.

Write it

Lead with the mission and the artifact wanted. Explain where the work stopped, what remains unresolved, and the next useful action or question. Give the recipient enough of the larger objective to understand why that step matters and what remains afterward.

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.

For implementation handoffs, direct the recipient to implement and verify the accepted outcome through completion, with independent adversarial-review checkpoints suited to the work. Preserve agreed checkpoints and choose useful ones from the task context; where the evidence is still missing, have the recipient choose and revise their placement as implementation develops. Place reviews where findings can change dependent work before it is built. Review the cumulative implementation and remaining plan, resolve findings, update the plan where those findings change it, and continue within the accepted scope. A checkpoint ends the review, not the assignment. Preserve explicit user limits and narrower research or review-only deliverables.

Carry forward unresolved findings and the task-start baseline so cumulative reviews can distinguish task changes from existing work. If implementation has not begun, have the recipient capture that baseline before editing. If an ongoing task lacks a baseline, state the gap rather than treating the handoff state as evidence of what predates the task.

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. Keep history when it explains the current direction or prevents repeating a failed approach. Omit transcript bulk and details the recipient can readily recover from the sources you identify.

The posture that makes it work

Say what you think is true and why. Distinguish explicit user decisions from observed behavior and your interpretation. The recipient should preserve the user's constraints while remaining free to re-check evidence, disagree with your conclusions, and choose its approach.

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.

Read the prompt as someone arriving cold: can they tell what to accomplish, what is known, and how to move forward without guessing at missing conversation? A thorough summary that leaves the next step unclear is not a finished handoff.

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.