菜单
Use when the user wants to continue in a fresh session, hand work to another agent, preserve current context near context limits, or create a compact continuation brief for later. Use when they say handoff, next session, fresh agent, continue tomorrow, context limit, or pass this to another agent.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when planning features, designing changes, stress-testing ideas, scoping work, or deciding how to approach non-trivial implementation. Use when the user says plan, design, shape, grill, think through, prototype the idea, or asks for an implementation plan.
Use when running an unusually strict code quality review focused on maintainability, structural simplification, abstraction quality, spaghetti growth, giant files, type boundaries, and whether a branch should be blocked despite passing tests. Use when the user says thermo-nuclear review, thermonuclear review, harsh maintainability review, deep code quality audit, or review this branch aggressively.
Use when creating, editing, pruning, splitting, merging, or testing agent skills. Use when deciding skill frontmatter, model-vs-user invocation, description triggers, supporting files, pressure scenarios, or whether a skill should exist.
Use when debugging any bug, failing test, flaky behavior, performance regression, unexpected runtime behavior, or reviewing an existing bug theory/postmortem. Use before reading source code to explain a symptom, before proposing a fix, or when tempted to patch from inspection.
Use when deciding which layer a piece of code belongs in (domain, application, or infrastructure), when business rules are mixed with framework/UI/database/HTTP code, when an external SDK or API response shape is leaking into core logic, when a test forces you to mock your own classes or sibling collaborators, or when unsure whether to introduce an interface/port/abstraction.
Use when implementing features, fixing bugs (including P0 hotfixes and production incidents), refactoring, removing dead code, or deleting legacy paths. Use when user mentions TDD, red-green-refactor, test-first, or when tempted to write code before tests. Use when fixing any bug — even trivial ones. Use when encountering or modifying existing code that lacks tests. Use when writing tests, deciding what to mock, or when tests are full of mocks, coupled to implementation, or break on every refactor.
| name | handoff |
| description | Use when the user wants to continue in a fresh session, hand work to another agent, preserve current context near context limits, or create a compact continuation brief for later. Use when they say handoff, next session, fresh agent, continue tomorrow, context limit, or pass this to another agent. |
Core principle: a handoff is a compact bridge to a fresh context window, not a new source of truth.
Write only what the next agent needs to continue. Reference durable artifacts instead of duplicating them.
Save the handoff in the OS temp directory, not the repo/workspace.
Use a clear filename:
/tmp/handoff-<topic>-<timestamp>.md
On macOS/Linux use $TMPDIR if set, otherwise /tmp. Do not create docs/.../handoff.md unless the user explicitly asks for a durable repo artifact.
Required sections:
# Handoff: <topic>
## Focus for next session
<what the next agent should accomplish>
## Current status
<what is done, what is in progress, what is not done>
## Durable artifacts to read
- <path or URL>: <why it matters>
## Decisions already made
- <decision + consequence>
## Next steps
1. <concrete action>
2. <concrete action>
## Open questions / risks
- <question or risk>
## Suggested skills
- <skill-name>: <why>
## Do not duplicate / do not touch
- <constraints, files, or decisions to avoid changing>
If the user gave a focus argument, tailor the handoff to that focus. Do not summarize the whole session if the next session only needs slice 2.
If a secret appeared in the conversation, redact it and add a security note: “A live-looking secret was present in prior context; do not copy it. Rotate/revoke if real.”
Recommend skills the next agent should load, based on the task:
| Next task | Suggested skills |
|---|---|
| continue design/planning | shaping |
| implement code | tdd, surgical-edits, relevant domain skill |
| debug a failure | systematic-debugging |
| decide code placement / mocking | placing-code-in-layers |
| edit skills | writing-skills |
| browser/UI work | frontend-design or browser automation skill if available |
Do not over-list. 1–4 suggested skills is enough.
Before finishing:
Example final response:
Wrote handoff: /tmp/handoff-import-wizard-slice-2-20260622-1530.md
Resume with: pi @/tmp/handoff-import-wizard-slice-2-20260622-1530.md "Continue from this handoff"