| name | handoff |
| description | Compress the live conversation into minimal, copy-paste-ready handoff prompt(s) — context referenced by file path, not pasted — so a fresh chat picks up cleanly. TRIGGER when: user says handoff, handover, new-chat / fresh-chat prompt, prompt to continue or resume elsewhere; user asks for separate prompts to explore N different things in their own chats. |
Handoff
Each prompt is self-contained, scoped to one thread, and points at the files holding the context instead of restating them.
When to use
- Context window is filling and the work should continue in a new chat.
- The chat has spawned several distinct threads the user wants to explore in separate chats — emit one prompt per thread.
- Wrapping a session and the user wants a clean resume later.
Skip when "resume the work in <doc>" already suffices — just hand them that path.
Protocol
Step 1 — Scope the handoffs
Read the user's request plus the conversation, then decide how many prompts and their boundaries:
- User named the items ("give 2 separate prompts — 1… 2…", "one for X, one for Y") → one prompt per item, in the order given. Note run-order if they're sequenced.
- One thread → one prompt.
- Ambiguous (chat sprawled across topics, user just said "handoff") →
AskUserQuestion: one combined prompt, or separate per thread — list the threads you'd split into so the choice is concrete.
Step 2 — Locate context to reference, don't restate
For each thread, point Read first: at the durable doc that holds the work's state — CONTEXT.md, plan.md, interview.md, a spec or discussion doc, else the specific source files, a commit, or a PR. In the prompt, carry only what that doc misses: current status, the last decision, the next action.
If no doc holds the state and the context won't fit a lean prompt, recommend creating one — default CONTEXT.md.
Step 3 — Draft each prompt
Fill the skeleton. Put each field's label on its own line, value(s) beneath; bullet the values when there's more than one. Omit empty fields — Status: only when resuming mid-work, not for a fresh-start handoff.
Repo:
<abs path>[ · Working dir: <dir> — never touch <x>]
Read first:
- <path>
- <path>
Status:
<one line — where things stand>
Goal:
<the one outcome this chat should reach>
Constraints:
- <read-only vs act intent — "analysis only, present for me to decide" vs "kick off the work">
- <don'ts; guardrails not guaranteed by the target's config>
Output:
<what to produce / the format expected>
- Don't restate the user's standing global preferences — the target session loads its own
CLAUDE.md. Carry only task-specific guardrails and the output contract.
Step 4 — Emit
Print each prompt as its own independently copy-pasteable block. If sequenced, add one run-order line after them ("Run 1 first; fire 2 once it lands").
Rules
- Minimal by default, context by reference. Point at bulk context by path; never paste it.
- Never merge distinct threads — one self-contained prompt each.
- Inline, not a file. The output is the pasted prompt; don't write it to disk unless asked.
- Reflect current state. Re-read the conversation's end state; drop any next-step that already happened.
- Externalize transient state. In-session task lists don't survive into a new chat — point the prompt at the durable doc as source of truth and tell it to rebuild its task list from there.
- Pin the intent. Every prompt says whether the new chat should act or only analyze.
