handover
Use when handing finished work over to code review — writing a PR description or packaging a change for review by a human, an agent, or both.
Install
npx skills add https://github.com/eai-org/agent-toolkit/tree/main/skills/handover
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install eai-org-agent-toolkit@llmmart
git clone https://github.com/eai-org/agent-toolkit.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole eai-org/agent-toolkit collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
Handover
Package a finished change so its reviewers never reconstruct intent from the diff. Write for a reviewer holding the diff and nothing else — no planning docs, no session, no knowledge that either exists.
Resolve the task
The input is any partial reference — ticket id, slug, a planning artifact's path — or nothing,
meaning the session's task; in a fresh session, infer it from branch name and recent commits.
Resolve it to the task's planning home: its <id>-<slug>/ directory (no ticket → no id) in
the project's planning directory, or its flat <id>-<slug>.* files in a shared family directory.
One clear match → proceed, stating the home and artifacts found. Inferred, or several candidates
(-vN, …) → confirm first. An explicit reference matching nothing local → stop and ask; never
guess.
Gather
Read these, skipping what doesn't exist:
- Task artifacts — every
<id>-<slug>*.mdin the planning home: ticket, requirements, plan, decisions log, prior review and handover rounds, whatever else matches. - Related tickets — one hop only: parent (climbing higher only past thin containers — a
parent that is only a title and a child list), direct predecessors/successors, explicit
relations, family-directory siblings; never expand their own relations. Local
.TICKET.mds first; fetch from the tracker only when a relation has no local file and looks load-bearing for a why — in doubt whether to fetch, ask. - The session, when it produced the change: decisions, pivots, constraints.
- The diff from the merge base with the target to
HEAD, plus commit subjects. Target: the one a prior review or handover round records, else the open PR's base, else the auto-detected default branch — of theupstreamremote when one exists, else oforigin, viagit ls-remote --symref <remote> HEAD; that failing (e.g. offline), the sole existing candidate amongmain,master,develop/development, preferringupstream's remote-tracking ref, thenorigin's, then the local branch; none or ambiguous → ask. State it andgit fetchits remote (local-only target → nothing to fetch; a failed fetch → ask, never diff stale refs). No merge base →git fetch --unshallowand retry, else ask — never diff against the target itself. An empty diff from the merge base → probably a wrong target: say so and ask, never match planned items against it. Before diffing, name uncommitted changes, untracked included, and ask which belong: the author commits those first; the rest stays out. - PR conventions — the repo's PR template, wherever the forge looks for one (e.g. root,
.github/,docs/; several → ask the author which), and any project skill for PR descriptions.
An earlier artifact — a ticket review, a prior round — counts only where nothing later settled the point.
Each source once, no deeper than the artifact needs: skim the diff whole; deep-read only the files you will name in the Review guide; for a visual, read just the function, component or directory it draws; never rebuild history commit by commit — which commit changed what is the diff's job, not yours.
Then match the plan's steps and acceptance criteria against the diff both ways — planned but absent, present but unplanned. Done when every source is read or confirmed absent and every planned item is matched against a non-empty diff.
Source every claim
State a "why", or any other claim about the change, only where a source gives it. Silence is not a source: never infer a fact from what the diff and docs leave unsaid ("untested so far", "nobody calls this"). A deviation nothing explains is asked of the author once; unanswered or unaskable, it ships flagged in plain words ("nothing records why — worth confirming"), since it may be an unintentional gap rather than a decision. Sourcing is your gate, not the reviewer's reading: it decides what you may write, and never appears in the text. When not sure, always ask. Never guess.
The artifact
<id>-<slug>.HANDOVER.md in the planning home, same base as the task's other artifacts; already
taken → <id>-<slug>.HANDOVER-2.md, -3, … — never overwrite. No planning home →
<id>-<slug>/ in the project's planning directory (none evident → .agents/plans/), id and slug
derived from branch name and changes, stating the choice. The file's first line is the PR title
in the repo's commit convention (# <title>); the body follows, paste-ready as the PR
description, and stands alone:
- Mention only what the reviewer can open — a tracker URL, or a file you verified is committed on the branch. Everything else — planning docs, decisions log, session, commit hashes — is neither linked nor named: write what it says ("this was meant to …"), never where it says it. A colleague's clarification is stated as what it settled, grounded in something openable ("the goal is X, which
- Related tickets go unmentioned unless a why depends on one ("built this way to prepare for
- Under a screen, ~400 words; the caps below are limits, not targets.
- A visual where the diff hides the change's shape — a flow across several files, a moved or
split layout, a reordered call chain — one per point, never a second for the same one. The
smallest shape that makes the point, from visuals.md, placed right under the line
it supports; it takes screen, not words. A visual is a claim: every node and
+/-line is in the diff or a source. - Never hard-wrap: forges (e.g. GitHub) render newlines in PR bodies as line breaks — one line per paragraph or bullet.
- Plain reviewer-facing wording, never this skill's vocabulary. Before drafting, actually invoke use-conversational-language — reciting its rules from memory does not count. Not installed: write plain prose instead, no dashes and no AI tells.
Sections, skipped only when truly empty:
- What and why — 2–3 lines.
- Decisions worth knowing — at most 5 lines, each: what was chosen or what departs from the plan, its why or the missing-why flag, and where in the code to see it.
- Review guide — the few files where judgment matters and why, what cannot be undone after merge among them; the rest named as mechanical.
- Known gaps — at most 3 shortcomings, assumptions or open questions the sources establish; none → omit the section, never manufacture one.
Repo PR template (if the repo has one) → the body is that template filled in: each section above goes into the template's matching section, else its description section; checkboxes and fields are ticked or filled only where a source backs them, else left blank; instructional comments stripped, fixed footers kept; the word cap counts only your prose. A project skill for PR descriptions overrides these rules where they differ. No template → the body is just the sections above.
Done when every section and template field is filled or knowingly skipped, every claim traces to a source, and nothing in the text points at something the reviewer cannot open.
Close
Print the file's project-relative path and say the body below the title is the PR description to paste; nothing else from the body. Done when both are printed.
Boundaries
- Modify no source files; the handover doc is the only file written.
- Read-only git plus
git fetch; never commit, push, or open/comment on a PR — publishing is the user's explicit call.
Files (agent-toolkit)
-
SKILL.md 7.7 KB
--- name: handover description: Use when handing finished work over to code review — writing a PR description or packaging a change for review by a human, an agent, or both. license: MIT metadata: version: "0.11" --- # Handover Package a finished change so its reviewers never reconstruct intent from the diff. Write for a reviewer holding the diff and nothing else — no planning docs, no session, no knowledge that either exists. ## Resolve the task The input is any partial reference — ticket id, slug, a planning artifact's path — or nothing, meaning the session's task; in a fresh session, infer it from branch name and recent commits. Resolve it to the task's **planning home**: its `<id>-<slug>/` directory (no ticket → no id) in the project's planning directory, or its flat `<id>-<slug>.*` files in a shared family directory. One clear match → proceed, stating the home and artifacts found. Inferred, or several candidates (`-vN`, …) → confirm first. An explicit reference matching nothing local → stop and ask; never guess. ## Gather Read these, skipping what doesn't exist: 1. **Task artifacts** — every `<id>-<slug>*.md` in the planning home: ticket, requirements, plan, decisions log, prior review and handover rounds, whatever else matches. 2. **Related tickets** — one hop only: parent (climbing higher only past thin containers — a parent that is only a title and a child list), direct predecessors/successors, explicit relations, family-directory siblings; never expand their own relations. Local `.TICKET.md`s first; fetch from the tracker only when a relation has no local file and looks load-bearing for a why — in doubt whether to fetch, ask. 3. **The session**, when it produced the change: decisions, pivots, constraints. 4. **The diff** from the merge base with the target to `HEAD`, plus commit subjects. Target: the one a prior review or handover round records, else the open PR's base, else the auto-detected default branch — of the `upstream` remote when one exists, else of `origin`, via `git ls-remote --symref <remote> HEAD`; that failing (e.g. offline), the sole existing candidate among `main`, `master`, `develop`/`development`, preferring `upstream`'s remote-tracking ref, then `origin`'s, then the local branch; none or ambiguous → ask. State it and `git fetch` its remote (local-only target → nothing to fetch; a failed fetch → ask, never diff stale refs). No merge base → `git fetch --unshallow` and retry, else ask — never diff against the target itself. An empty diff from the merge base → probably a wrong target: say so and ask, never match planned items against it. Before diffing, name uncommitted changes, untracked included, and ask which belong: the author commits those first; the rest stays out. 5. **PR conventions** — the repo's PR template, wherever the forge looks for one (e.g. root, `.github/`, `docs/`; several → ask the author which), and any project skill for PR descriptions. An earlier artifact — a ticket review, a prior round — counts only where nothing later settled the point. Each source once, no deeper than the artifact needs: skim the diff whole; deep-read only the files you will name in the *Review guide*; for a visual, read just the function, component or directory it draws; never rebuild history commit by commit — which commit changed what is the diff's job, not yours. Then match the plan's steps and acceptance criteria against the diff both ways — planned but absent, present but unplanned. Done when every source is read or confirmed absent and every planned item is matched against a non-empty diff. ## Source every claim State a "why", or any other claim about the change, only where a source gives it. Silence is not a source: never infer a fact from what the diff and docs leave unsaid ("untested so far", "nobody calls this"). A deviation nothing explains is asked of the author once; unanswered or unaskable, it ships flagged in plain words ("nothing records why — worth confirming"), since it may be an unintentional gap rather than a decision. Sourcing is your gate, not the reviewer's reading: it decides what you may write, and never appears in the text. When not sure, always ask. Never guess. ## The artifact `<id>-<slug>.HANDOVER.md` in the planning home, same base as the task's other artifacts; already taken → `<id>-<slug>.HANDOVER-2.md`, `-3`, … — never overwrite. No planning home → `<id>-<slug>/` in the project's planning directory (none evident → `.agents/plans/`), id and slug derived from branch name and changes, stating the choice. The file's first line is the PR title in the repo's commit convention (`# <title>`); the body follows, paste-ready as the PR description, and stands alone: - **Mention only what the reviewer can open** — a tracker URL, or a file you verified is committed on the branch. Everything else — planning docs, decisions log, session, commit hashes — is neither linked nor named: write what it says ("this was meant to …"), never where it says it. A colleague's clarification is stated as what it settled, grounded in something openable ("the goal is X, which <ticket URL> needs"); naming the colleague is the last resort. - **Related tickets go unmentioned** unless a why depends on one ("built this way to prepare for <X>") — then cite its tracker URL. - **Under a screen**, ~400 words; the caps below are limits, not targets. - **A visual where the diff hides the change's shape** — a flow across several files, a moved or split layout, a reordered call chain — one per point, never a second for the same one. The smallest shape that makes the point, from [visuals.md](visuals.md), placed right under the line it supports; it takes screen, not words. A visual is a claim: every node and `+`/`-` line is in the diff or a source. - **Never hard-wrap**: forges (e.g. GitHub) render newlines in PR bodies as line breaks — one line per paragraph or bullet. - **Plain reviewer-facing wording**, never this skill's vocabulary. Before drafting, actually invoke use-conversational-language — reciting its rules from memory does not count. Not installed: write plain prose instead, no dashes and no AI tells. Sections, skipped only when truly empty: 1. **What and why** — 2–3 lines. 2. **Decisions worth knowing** — at most 5 lines, each: what was chosen or what departs from the plan, its why or the missing-why flag, and where in the code to see it. 3. **Review guide** — the few files where judgment matters and why, what cannot be undone after merge among them; the rest named as mechanical. 4. **Known gaps** — at most 3 shortcomings, assumptions or open questions the sources establish; none → omit the section, never manufacture one. **Repo PR template** (if the repo has one) → the body is that template filled in: each section above goes into the template's matching section, else its description section; checkboxes and fields are ticked or filled only where a source backs them, else left blank; instructional comments stripped, fixed footers kept; the word cap counts only your prose. A project skill for PR descriptions overrides these rules where they differ. No template → the body is just the sections above. Done when every section and template field is filled or knowingly skipped, every claim traces to a source, and nothing in the text points at something the reviewer cannot open. ## Close Print the file's project-relative path and say the body below the title is the PR description to paste; nothing else from the body. Done when both are printed. ## Boundaries - Modify no source files; the handover doc is the only file written. - Read-only git plus `git fetch`; never commit, push, or open/comment on a PR — publishing is the user's explicit call. -
visuals.md 1.5 KB
Loading…
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.