handoff-prompt-writer

Handoff Prompt Writer

Create one self-contained prompt file that a successor agent can execute. The default output is handoff-prompt.md in the active project root. Write the file; do not stop at showing a draft in chat.

Quick Start

1. Confirm the handoff scope from the current request. Use the user-specified output path, or default to handoff-prompt.md. For a multi-repo task, place it at the nearest common workspace root and label every repository.
2. Gather the smallest authoritative evidence set:
   • current user objective, intent, constraints, and explicit decisions;
   • relevant plans, progress notes, results, specifications, and logs;
   • live repository state, changed files, branch and revision, and diffs;
   • test, build, render, or command results that support completion claims;
   • unresolved questions, blockers, risks, and remaining work.
3. Reconcile conflicts. Current explicit user direction controls intent and scope. Live files, version-control state, and fresh tool output control implementation status. Treat older notes and conversation claims as context, not proof, when they disagree with inspectable state.
4. Fill assets/handoff-prompt.template.md. Replace every placeholder, remove empty optional rows, and keep the result addressed directly to the successor.
5. Point to exact files, commands, commits, and evidence instead of pasting long source material or a transcript.
6. Remove secrets, tokens, credentials, personal data, private reasoning, and irrelevant history. Preserve material caveats and authority boundaries.
7. Re-read the file cold. Verify that its first actions are executable and that every completed claim has inspectable evidence.

Read references/source-notes.md only when maintaining the handoff contract or adapting it to a new agent runtime.

Prompt Contract

The generated file must contain:

• a direct successor role and operating contract;
• the primary objective and observable definition of done;
• intent and background that explain why the work exists;
• in-scope, out-of-scope, and confirmation-required actions;
• current status, repository or artifact state, and important decisions;
• completed work paired with files, commands, tests, or other evidence;
• an artifact map with paths and why each item matters;
• remaining work ordered by dependency and impact;
• blockers, unknowns, risks, and what would resolve them;
• the first one to three executable actions;
• verification requirements and final delivery expectations.

Use verified, inferred, and unknown labels only where ambiguity matters. Do not burden obvious facts with labels. A successor should know which claims it can trust and which it must check.

Evidence and Source Priority

Use this order for conflicts:

1. Current explicit user instruction for goal, scope, and authority.
2. Current filesystem, repository, external-system state, and fresh tool output.
3. Accepted specifications and maintained project artifacts.
4. Plans, progress notes, prior summaries, and conversation memory.

Do not infer intent from a diff when the user stated it directly. Do not claim that work is complete because a plan says so or a file exists. State what was not checked.

Continuation Behavior

Write the handoff as an instruction to continue, not as a retrospective report. The successor should be told to:

• re-ground on the named evidence before editing;
• preserve verified work and avoid repeating completed investigation;
• continue in scope without asking for routine reversible actions;
• pause only for destructive or irreversible actions, material scope changes, or information only the user can supply;
• validate outcomes before reporting them;
• update or replace stale handoff information if another transfer is needed.

If an existing handoff prompt is present, read it first. Preserve still-valid context and user decisions, but replace stale status and evidence rather than appending another conflicting summary.

Output Contract

Return the path to the generated prompt, its overall continuation status, and the first action encoded in it. Mention evidence gaps only when they affect the successor's ability to proceed. The file itself is the deliverable.

Gotchas

• Do not dump the conversation. Preserve decisions and consequences, not turns.
• Do not include hidden or private reasoning. Record concise rationale and evidence that another agent can inspect.
• Do not bury the next action after background. Make the start path explicit.
• Do not copy secrets from environment files, logs, authentication output, or tool results into the prompt.
• Do not silently combine unrelated repositories or tasks into one handoff.
• Do not use this skill when the user needs the stricter handoff.md plus context-pack.json parity contract.