本文へ移動
cccskills
無料GitHub で公開

ai-session-handoff-writer

当一个长任务要跨 session 继续、或者要交给另一个人、另一个 Agent 接手时使用。把当前进度压成一份可移植的交接文档:目标、已确认完成的事、关键决定及其理由、证据与验证状态、遗留风险、下一步的具体动作、重启须知。事实与推断分开标注,没验证过的绝不写成已完成。

写的是跨 session 的交接,不是子代理交回主代理的那种短单。

インストール方法を見る

含まれるファイル(2)

  • SKILL.md7.6 KB
  • handoff-template.md994 B

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

ai-session-handoff-writer

What this skill does

This skill turns an in-progress task into a handoff document that another agent or person can continue with minimal context loss.

It is optimized for long-running work, session resets, model switches, ownership changes, and paused tasks.

Use this skill when

  • A task will continue in another session
  • Work is being transferred between people or agents
  • A project has multiple open threads and needs a clean restart point
  • You need to preserve not just status, but also why decisions were made

Do not use this skill when

  • A one-line chat update is enough
  • There is no meaningful progress, decision, or artifact to hand off
  • The recipient is not allowed to see the underlying work or data
  • The main need is a code diff review rather than a task-level handoff

Required inputs

Collect or infer these before writing:

  1. Task objective — what the work is trying to achieve
  2. Current state — what is already done and what remains open
  3. Artifacts — files, documents, links, tickets, datasets, or outputs that matter
  4. Key decisions — important choices already made
  5. Rationale — why those choices were made
  6. Verification status — what has been checked versus assumed
  7. Risks or blockers — what could stop or mislead the next worker
  8. Next step — the best immediate continuation action

If any required input is missing, keep the section and label it clearly as [missing] rather than inventing details.

Evidence labels

Use these labels consistently inside the handoff:

  • [confirmed] — directly verified from files, tests, outputs, or trusted records
  • [inferred] — strongly supported but not directly verified in the current context
  • [unverified] — reported, assumed, or still needs checking
  • [missing] — required information is not currently available

Do not blur these categories.

Workflow

1) Define the handoff target

State who or what should be able to continue after reading the handoff:

  • another general-purpose agent
  • a coding-focused agent
  • a human collaborator
  • your future self in a later session

Write for the least-informed plausible recipient.

2) Separate facts from interpretation

Before drafting, sort notes into:

  • completed work
  • decisions
  • evidence
  • open questions
  • next actions

If a claim cannot be supported, mark it [inferred] or [unverified].

3) Capture only the continuation-critical context

Prefer information that changes what the next worker should do:

  • exact file paths or artifact names
  • commands already tried and their outcomes
  • constraints, approvals, or boundaries
  • rejected approaches worth not repeating
  • ordering dependencies

Avoid dumping long transcripts unless they contain irreplaceable reasoning.

4) Write the handoff in the standard structure

Use the template in handoff-template.md.

Minimum required sections:

  1. Objective
  2. Current status
  3. Completed work
  4. Key decisions and rationale
  5. Evidence and verification
  6. Outstanding issues and risks
  7. Recommended next actions
  8. Restart notes

Optional, as in the template:

  1. Open questions: include it when something is waiting for a decision from the user or the next owner; leave it out when nothing is.

5) Validate before sending

A handoff is only complete if the next worker can answer all of these:

  • What is the goal?
  • What is already done?
  • What should not be undone lightly?
  • What still needs verification?
  • What is the very next action?
  • Where are the relevant artifacts?

If any answer is unclear, revise the handoff.

Output rules

  • Prefer bullets over long prose
  • Put the next action near the end as a numbered list
  • Name files and paths explicitly
  • Preserve uncertainty labels
  • Keep tone neutral and operational
  • Never write the value of a secret (API key, token, password, connection string, private key) into the handoff, whoever the recipient is and even if the user says it is fine: handoffs get forwarded and pasted into other sessions. Write where the value lives and its variable name instead (for example DB_URL in .env), and tell the recipient to obtain it themselves
  • Redact other private data unless the recipient is explicitly authorized to see it

Quality checklist

Before finalizing, confirm:

  • All required sections are present
  • Every major claim has an evidence label
  • Paths, URLs, or artifact identifiers are explicit where relevant
  • Decisions include reasons, not just outcomes
  • Risks mention likely failure modes or missing information
  • Next actions are ordered and concrete
  • The document contains no secret values, only where they live and their variable names
  • Other private data appears only if the recipient is authorized to see it

Failure modes to avoid

  • Writing a status summary with no restart path
  • Mixing verified facts with guesses
  • Omitting the rationale behind important choices
  • Burying blockers in narrative text
  • Providing ten possible next steps instead of one recommended sequence
  • Repeating raw chat history instead of distilling it

Handoff compression guidance

When the context is large, compress in this order:

  1. Keep decisions and rationale
  2. Keep artifact locations
  3. Keep verification results
  4. Keep blockers and dependencies
  5. Summarize routine background detail last

Suggested prompt pattern

When using this skill, produce a handoff document that:

  • preserves verified progress
  • distinguishes confirmed vs inferred vs unverified information
  • lists concrete artifacts and locations
  • explains key decisions and why they were made
  • ends with a short ordered next-step plan

Deliverable

Return a completed handoff document based on this skill's own handoff-template.md.

Not the same as the repository's short handoff

templates/handoff-template.md at the repository root is a different artifact: a ~30-line note from a sub-agent back to its parent, inside one session, carrying only the result, the evidence location and the next step. It assumes the reader already has the context.

This skill writes the cross-session handoff, where the next reader has no context at all, so it additionally carries decisions with their rationale, what was already tried and should not be repeated, and restart notes.

Pick by one question: does whoever picks this up already have this session's context? Yes → use the repository's short template. No → use this one.

变更记录 / Changelog

版本日期变更类型
0.1.12026-09-26必备节数与模板对齐:八节必备,第九节"待确认问题"可选(用例原先要求九节齐全);密钥一律不写值,删掉"明确授权后可写入"的例外,Quality checklist 同步改,模板的重启须知加"凭据只写位置和变量名"patch
0.1.02026-08-30初始版本minor

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

当学生说"Common App 活动栏写不下"、"这段活动描述超字数了"、"帮我把这几条压到限制内"时使用。先按官方申请表核对当季字符限额,再逐条检查;一次把所有条目量完并直接给出压缩版,只在需要砍掉实质内容时才用选择题问该保哪部分。每改一版都用工具重新数一遍字符,绝不凭感觉说"这样应该够了"。只压缩,不代写空白。

日本語の概要は準備中です。原文の説明を表示しています。

ChromiteCr/Skills42026年10月16日 更新

当学生说"我参加过一些活动但没整理过"、"帮我把经历录进去"、"我的活动档案是空的"时使用。让学生一口气把记得的活动说出来,直接把这段话拆成结构化的长期事项(角色、组织、起止时间、做了什么、结果、级别),说不清的字段留空标"待补"而不是编,再用一张选择题卡把级别判定与是否重复这类关键项一次问完,然后提案入库。只记录学生真正说出口的内容,级别按事实判定,绝不往高了写。

日本語の概要は準備中です。原文の説明を表示しています。

ChromiteCr/Skills42026年10月16日 更新

当学生问"招生官会怎么看我的档案"、"我的背景还差什么"、"帮我评估一下现在的申请竞争力"时使用。以顶尖大学招生官的视角通读学生的档案、长期活动与已写下的反思,不问任何前置问题直接产出一份指名道姓的坦诚点评:这份材料的亮点、招生官真实的解读方式、明显的短板与空白、以及接下来最值得做的一到两件事。只读不写,不修改学生的任何数据,也不代写任何申请材料。

日本語の概要は準備中です。原文の説明を表示しています。

ChromiteCr/Skills42026年10月16日 更新

当使用者说"帮我填词"、"按这个字数模板写歌词"、"用多Agent对抗式写词"、"这首副歌帮我想几句"时使用,前提是字数骨架和这首歌写什么都已经有了;只有主题、还没定写什么时先用 lyric-concept-builder。拿一份字数骨架和意象系统,派三个 sonnet 子代理按不同路子并行独立起草(场景派 / 口语派 / 意象派),再派一个红队子代理逐句挑毛病,主 Agent 按句裁决拼稿,最后统一辙口与意象。收敛判据卡死:字数用 lyric-doctor 的脚本对模板数、每段同辙、hook 不含陈词、主歌二对主歌一有推进。不抄现有歌词,不替使用者决定这首歌要说什么。

日本語の概要は準備中です。原文の説明を表示しています。

ChromiteCr/Skills42026年10月16日 更新

当想知道一段 AI 回答"哪些能直接用、哪些必须先验证、哪些只能当建议"时使用。按声明类型分四级——低风险草稿材料、观点与策略建议、可查证的事实声明、会改变状态的可执行指令——每级配一句默认处置动作;再按影响半径排序,给出最小的下一步检查。可执行类一律标"不能直接执行;先做代码审阅/沙箱测试/回滚准备",其中破坏性或提权的命令排在最前。教的是一套可复用的复查习惯,不是替使用者把所有内容都核实一遍:本 skill 只分级、排序、定处置,要逐条核实走 ai-output-fact-checker,整段代码要做入库体检走 ai-code-onboarding-checklist。

日本語の概要は準備中です。原文の説明を表示しています。

ChromiteCr/Skills42026年10月16日 更新

当拿到一段 AI 生成的代码,还没决定要不要信、要不要跑、要不要往上继续加东西时使用。先做一遍客观体检:行数与嵌套深度、有没有测试、有没有硬编码密钥或遗留 TODO、依赖是不是真实存在、有没有危险调用(eval/exec、shell=True、递归删除、反序列化、提权、外发网络,按正文给的 grep 模式逐类查);再给风险评级和下一步动作。老实写明它查不了算法正确性——体检通过不等于代码是对的。回答里文字、事实和代码混在一起时先用 ai-answer-triage 分级;只审相对已有代码的一份改动用 ai-diff-review-protocol。

日本語の概要は準備中です。原文の説明を表示しています。

ChromiteCr/Skills42026年10月16日 更新

ChromiteCr のスキルをすべて見る

このスキルの問題を報告する