Claude Code Handoff
Write one copy-paste prompt for Claude Code. It must be cold-start: Claude can continue without this thread.
Default output: return only the prompt. If the user is designing or reviewing the handoff itself, a short note before the prompt is allowed.
One Sentence
A handoff prompt gives Claude context, starting points, proof targets, and a strong bias to use /codex:rescue creatively, while leaving Claude free to rethink the plan and choose the delegation shape.
Mental Model
Claude is the continuing agent:
Claude owns:
orchestration
context selection
ambiguity and tradeoffs
implementation direction
`/codex:rescue` delegation choices
final synthesis
Codex can help with:
broad search and grep
diff excavation
exact file reads
focused edits
command execution
tests and typechecks
browser checks
cleanup
adversarial checks
small prototypes
Do not pre-orchestrate the session. Prime Claude to use the literal /codex:rescue command, suggest seams when they are obvious, and explicitly allow Claude to revise, split, skip, or invent Codex calls after reading live context.
Ground Before Writing
Gather the context the next agent should not have to rediscover.
For coding handoffs, usually collect:
git status --short --branch
git diff --name-status
git diff -- <relevant paths>
Also read the key files, tests, specs, ADRs, PR notes, or logs that ground the prompt. Include commands already run and their pass/fail status when useful.
For non-coding handoffs, gather the equivalent source material: docs, notes, links, current conclusions, open questions, and the artifact the next agent should produce.
Skip heavy grounding only for obviously tiny handoffs or when the user asks for a fast conversational prompt.
Prompt Contents
Include only live context. A small handoff should fit in 12 to 20 dense lines; a large handoff may need short sections.
Cover what helps Claude start:
Mission:
One outcome and the artifact wanted.
State:
Branch, dirty worktree, staged files, checks run, and what remains open.
Source paths:
Exact files, diffs, commands, PRs, specs, ADRs, or notes to inspect first.
Current read:
What seems true, recommendations so far, and why. Avoid calling these settled unless an ADR, code path, or explicit user decision makes them settled.
Open questions:
What Claude should decide or re-check. Invite challenge when code, tests, ADRs, or product constraints contradict the current read.
Watch-outs:
Only real hazards: dirty user work, destructive git, production or deploy actions, migrations, security, licensing/package boundaries, obsolete paths to avoid, or explicit user non-goals.
Codex posture:
Tell Claude to look actively for work worth delegating through `/codex:rescue`. For substantial work, list candidate seams, but make them examples, not a queue.
Proof and stop:
Likely verification commands or evidence targets, plus where to stop.
Codex Posture
Always prime Claude that Codex is available through the literal /codex:rescue command. For tiny handoffs, one sentence is enough:
Use /codex:rescue where it buys speed, breadth, focused execution, or independent verification.
For substantial coding, review, debugging, migration, or verification work, suggest candidate /codex:rescue seams. Name the command in the handoff so Claude does not have to infer it from “Codex.” Shape seams as options:
Candidate Codex seams:
- Diff excavation: inspect <range/paths>; return keep/rewrite/drop map with risks.
- Focused implementation: edit <bounded paths> for <one wave>; return changed files, diff summary, risks, and verification.
- Verification: run <commands>; return pass/fail output and likely next fix.
Rules for candidate Codex seams:
- one job, bounded sources, one deliverable, clear stop condition
- exact inputs when known
- commit-sized waves for implementation
- disposable worktree or clearly named branch when edits are non-trivial
- result packets include changed files, diffs, command output, risks, and recommendations
- Claude may revise, split, skip, or add jobs after reading live context
What To Avoid
Do not turn the handoff into:
an execution script
an unchallengeable decision record
a mandatory Codex job queue
a rigid constraints list
a duplicate spec or pasted diff dump
a transcript summary with no source paths
Prefer current read over settled decision. Prefer watch-out over prohibition. Prefer candidate seam over commandment.
Stop Condition
The recipient should know what counts as done:
Stop after:
a review memo
a PR-ready diff
a clean implementation branch
a verified command output set
a blocker list with the smallest remaining decisions
For a /goal one-liner, use agent-goal. For a human-facing progress summary, use progress-summary.