claude-handoff

Claude Handoff

Orchestrates read-only investigation or implementation within the current session after explicit plan approval. Task-handoff instead writes a decision-complete file for a fresh, separate session — use it for work continuing later or elsewhere; use an in-session handoff to implement an approved plan now.

If these instructions are already present from a slash or dollar invocation, follow them directly; do not invoke this skill again through a skill tool.

Follow the Contract, then use Claude Code's in-session Agent workflow to research during planning and implement the approved plan.

Contract

• Run only after explicit user invocation. Classify a task research-only when its requested outcome is findings, evidence, or an assessment, with no repository changes or implementation plan requested. All handoffs may run in any host mode; implementation handoffs must pass through the Plan Phase and receive explicit user approval before launch.
• Claude owns decisions, the final plan, and orchestration. Delegate investigation to read-only research subagents only when task and repository evidence make it useful before planning.
• Research agents gather evidence and report findings; never edit files, decide design, or return plans.
• Implementation agents implement their assigned part of the approved plan — inspect, edit, validate — never redesign or return another plan.
• Use the smallest effective team (one agent is valid); add agents only when decomposition materially improves latency, correctness, or verification. Never exceed eight implementation agents per handoff.
• Size every brief before finalizing the team: estimate wall-clock time and split anything likely to exceed roughly 25-30 minutes into parallel disjoint scopes or dependency waves, adding an integration agent as needed.
• Use at most three research agents, stable IDs R1-R3, counted separately from the eight implementation agents.
• Keep Claude's own work to orchestration, integrity checks, failure handling, and conditional polish passes.
• Let Claude Code choose foreground or background delivery; never pass run_in_background to the Agent tool. Only a direct terminal result or error, or a later completion or failure notification, settles an agent — a launch acknowledgement, native task row, quiet period, or surfaced permission prompt does not. Reconcile a wave only after every required terminal outcome arrives.
• Treat an explicit user model preference (e.g. Sonnet, Opus) as an orchestration constraint on every research and implementation agent unless scoped narrower — never substitute complexity-based selection. If the Agent tool cannot launch that model, report the incompatibility and ask before falling back.
• Treat the approved outcome, not the initial manifest or its write scopes, as the authorization boundary: when implementation reveals a related in-repository fix or evidence change the outcome requires, Claude may extend the handoff and launch follow-on agents without asking again. The discovering subagent still stops at its assigned scope and returns evidence; Claude owns scope expansion, coordination, and delegation.