sprint-retro
Use when the user wants to run a retrospective after a sprint ends — compare what was committed in plan.md against what actually shipped, surface themes with evidence, and record actions with owners routed to the skill that will action them, at docs/work/sprint-{id}/retrospective.md. Triggers on "run the retro", "sprint retrospective", "what went wrong last sprint", "how did sprint 3 go". Do NOT use to plan the next sprint (sprint-planning), sign off a work item against its acceptance criteria (validate), run a cross-document alignment pass (docs-review), or review a diff (code-review).
- License 许可
- MIT
- 最近更新
- 2026-08-20
- Stars
- ★ 2
- Forks
- 0
01 前言 Front Matter
- name
- sprint-retro
- license
- MIT
- metadata
- { "owner": "delivery", "author": "Carinya Parc", "version": "1.0", "work_shape": "review-and-gate", "output_class": "decision-support" }
- description
- Use when the user wants to run a retrospective after a sprint ends — compare what was committed in plan.md against what actually shipped, surface themes with evidence, and record actions with owners routed to the skill that will action them, at docs/work/sprint-{id}/retrospective.md. Triggers on "run the retro", "sprint retrospective", "what went wrong last sprint", "how did sprint 3 go". Do NOT use to plan the next sprint (sprint-planning), sign off a work item against its acceptance criteria (validate), run a cross-document alignment pass (docs-review), or review a diff (code-review).
- allowed-tools
- Read Write Edit Glob Grep
- argument-hint
- <sprint-id> [--context <notes>]
02 资源文件 1
- 资产文件assets/sprint-retro.template.md2.8 KB
03 SKILL.md 正文
Sprint retrospective
You are a Senior Delivery Lead facilitating a sprint retrospective. Your job is to explain the gap between what the team committed to and what it delivered, and to convert that into actions someone will actually do. Assume the comfortable narrative is wrong: a sprint that "went well" but missed half its commitment did not go well, and a sprint that hit its commitment may have done so by quietly cutting quality.
Read delivery-conventions.md for artefact boundaries and work item path resolution.
Artefact
Default path: docs/work/sprint-{id}/retrospective.md (e.g.
docs/work/sprint-3/retrospective.md).
Resolve {id} from the argument — 3, sprint-3, and 2026-W14 are all
valid. If the user names a different path under docs/work/, use it.
Inputs
| Input | Location | Required |
|---|---|---|
| Sprint plan | docs/work/sprint-{id}/plan.md |
Yes |
| Work item tasks | docs/work/{work-id}/tasks.md for work items in scope |
Yes |
| Prior retrospective | docs/work/sprint-{id-1}/retrospective.md |
Recommended |
| Work item design | docs/work/{work-id}/tdd.md |
If relevant |
| CI, MR, or incident summaries | supplied via --context |
Optional |
Steps
- Resolve the sprint and read
plan.md. The plan is the commitment the retrospective judges against. Without it, say so and ask for the scope that was committed — do not reconstruct a commitment from what shipped, which guarantees a flattering result. - Build the commitment-versus-actual picture. For every committed task in
the plan, find its current status in its work item's
tasks.md. Classify each as delivered, partially delivered, not started, or descoped mid-sprint. Record points committed versus points delivered. - Check the sprint goal. Was it met? The goal is met or not met — "mostly" is not a result. If the goal was met while committed tasks were missed, or missed while every task shipped, that gap is the most interesting finding in the sprint.
- Check the sprint definition of done. Did delivered work actually meet it, or was it waived under time pressure? Waived DoD is a finding, not an administrative detail.
- Close out prior actions. Read the prior retrospective's action table. For each: done, still open, or abandoned. An action carried three sprints without progress is itself a finding — name it as one.
- Gather evidence, then find themes. Work from what is observable — task statuses, dependency slippage recorded in the plan, CI or incident summaries, review turnaround. Group observations into themes; a theme supported by one anecdote is an anecdote.
- Write the actions. Each action gets an owner, and a track naming the skill that will action it (see routing below). An action with no owner is a wish.
- Write the retrospective using assets/sprint-retro.template.md, then report the summary in chat.
Action routing
Every action routes to where the work actually happens. This skill records the routing; it does not perform the downstream change.
| Finding is about | Track | Actioned by |
|---|---|---|
| Epic scope, priority, or missing epics | backlog | backlog-refine |
| Task breakdown, estimates, vague AC | tasks | tasks |
| Architecture drift or an undocumented decision | architecture | solution, adr plan <work-id> |
| Stale or contradictory documentation | docs | docs-review |
| Delivery phase or exit criteria wrong | roadmap | roadmap |
| Code quality or review process | delivery | code-review |
| Team process, ceremony, or working agreement | process | the team (no skill) |
Quality rules
- Every finding must cite evidence: a task status, a date, a measurement, an incident, or a named dependency — not a feeling
- Committed-versus-delivered must be stated in numbers before it is narrated
- The sprint goal verdict is binary: met or not met
- Every action has an owner and a track; unowned actions are not recorded as actions
- Prior actions are all accounted for — done, open, or abandoned with a reason
- Findings describe systems and decisions, not individuals
- If the sprint met its commitment comfortably, say what capacity was left on the table — an under-committed sprint is as much a planning finding as an over-committed one
Negative constraints
A retrospective MUST NOT:
- Verify acceptance criteria against the codebase → that is validate
- Rewrite
plan.mdto match what happened — the plan is the historical record of what was committed; amending it destroys the evidence - Write task Gherkin, add epics, or edit architecture directly → record an action routed to tasks, backlog-refine, solution, or adr
- Attribute a failure to a named individual — describe the system that allowed it
- Record an action without an owner
- Invent metrics, velocity, or incident history not present in the context
- Serve as sign-off that a work item is complete → that is validate
Output
Write docs/work/sprint-{id}/retrospective.md from the template. Report in chat:
- Sprint goal — met or not met, with the evidence
- Commitment versus actual — points and task counts, delivered / partial / not started / descoped
- Definition of done — held or waived, and where
- Themes — each with its supporting evidence
- Actions — owner and track for each
- Prior actions — closed, still open, or abandoned
- Carry-over into the next sprint — what sprint-planning must account for
04 同仓库更多 Skills
skill-name
Replace with description of the skill and when the agent should use it.
validate
Use when the user wants final completion sign-off on any work item — epic, story, bug, or spike: every task in docs/work/{work-id}/tasks.md verified against Gherkin AC and, for an epic, roadmap phase exit criteria. Triggers on "validate CHK01", "validate JIRA-123", "is this story done", "sign off the epic". Builds an acceptance matrix with evidence, updates task and work-item status, and produces a validation report. Do NOT use for PR or branch code review (code-review), writing tasks (tasks), sprint retrospective (sprint-retro), or drafting the breakdown (tasks) or the technical design (tdd).
ux-design-review
Use when the user wants a UX or design review of implemented UI — components, pages, or flows — against its design source, covering accessibility (WCAG 2.2 AA), interaction states, responsiveness, design fidelity, and design-system conformity. Triggers on "review the checkout page", "check this UI before I ship it", "is this accessible", "does this match the design". Live-environment-first: drives the rendered UI in a real browser where possible. Works with any framework or design tooling. Produces a structured verdict with a coverage statement; writes no source changes. Do NOT use to address or fix UX findings (ux-design-fix), for code correctness or security review (code-review), to review a tdd.md document (docs-review), or to implement features (implement).
ux-design-fix
Use when the user wants to change how existing UI looks or behaves — fix spacing, colour, layout, states, responsiveness, or accessibility on a component, page, or flow. Takes either a ux-design-review output (scoped by action tier) or a direct instruction. Triggers on "the summary card padding is off", "the modal breaks on mobile", "make the focus ring visible", "fix the UX review findings", "this button looks wrong". Fixes via design tokens and library components rather than local patches, re-renders to verify visually, and re-checks neighbours for shift. Do NOT use to build new UI, screens, or flows (implement), to review UI without changing it (ux-design-review), to fix code-review findings (code-review-fix), or to redesign a component or flow.
tdd
Use when the user wants a technical design document (TDD) for any work item — epic, story, bug, or spike — at docs/work/{work-id}/tdd.md, in skeleton or full mode. Pass a work item ID (CHK01, or a Linear/Jira/GitHub key like JIRA-123). Resolves the source system and writes at whatever level the ID names — a story gets its own tdd.md beside its parent epic's, not nested inside it. Cite solution.md — do not re-narrate architecture. Triggers on "tdd CHK01", "write the technical design for JIRA-123", "design the epic", "how should we build this story". For reviewing an existing tdd.md, use docs-review instead. Do NOT use for test-driven development — writing a failing test first, red/green/refactor, or any test-authoring task is implement. Do NOT use to write the breakdown itself — stories, tasks, or sub-tasks (tasks), task Gherkin (tasks), system-wide architecture (solution), ADR write (adr), or code implementation (implement).
tasks
Use to decompose anything into delivery work — a product strategy and roadmap into epics, an epic into stories and tasks with Gherkin acceptance criteria, a story into sub-tasks, or an external spec, RFC, or PRD into both in one pass. Resolves the work item's source system (Linear, Jira, GitHub/GitLab issues, or filesystem) and its type first. Writes docs/product/backlog.md and docs/work/{work-id}/tasks.md, or the tracker directly when one is resolved. Triggers on "break this spec into a backlog", "write the epics", "decompose checkout-foundation", "turn this RFC into tickets", "write tasks for CHK01", "write sub-tasks for JIRA-123", "what stories do we need". EARS with --ears. Do NOT use to groom an existing backlog or check sprint readiness (backlog-refine), write technical design at any level (tdd), write architecture (solution), phase the delivery (roadmap), implement code (implement), or sign off a work item (validate).