An experiment in applying Scrum principles with AI teammates. A human remains accountable for product decisions.
A project-neutral skill for Codex and Claude Code. Invoke it to create or resume a Scrum workspace, refine a backlog, plan and deliver a bounded slice, inspect outcomes, improve the process and leave a reliable handoff.
The human Product Owner can work in any domain. Engineering, QA, design, security and domain reviewers are selected for the project. Legal counsel, specific models, frontend packages and vendor plugins are optional, not built-in requirements.
Version 0.5.1 · MIT · Two skills: scrum-skill and scrum-dashboard. No bundled MCP server, lifecycle hook, telemetry or automatic cross-project synchronization. The optional workspace helper uses Python 3.10+ and the standard library.
Run the inventory example to see real scaffold output. The image summarizes that helper output; it is not a native-host execution screenshot.
Release and marketplace submission guide explains the plugin, standalone skills and directory routes.
Choose one installation route to avoid duplicate skill names.
codex plugin marketplace add kevanwee/scrum-skill --ref v0.5.1
codex plugin add scrum-skill@scrum-skillsReload/start a session if needed, then invoke:
$scrum-skill Set up Scrum for this project. I am the Product Owner.
Our goal is a small inventory app. Use engineering, QA and design agents
where available. Create the artifacts and propose the first usable slice.
The repo catalog is a distribution source, not an official OpenAI directory listing. Host/plugin availability varies by surface. See official package/catalog guidance.
claude plugin marketplace add kevanwee/scrum-skill
claude plugin install scrum-skill@scrum-skillsReload/start a session if needed. Plugin skills use a namespace:
/scrum-skill:scrum-skill Set up Scrum for this project and propose a first Sprint.
For local development after cloning, claude --plugin-dir /path/to/scrum-skill loads the plugin for that session. See Claude plugin guidance for supported surfaces, scopes and trust settings.
Install just skills/scrum-skill/ if you do not need a plugin catalog. For Codex, ask $skill-installer to install that path from this repository at v0.5.1, or copy it into your project's .agents/skills/scrum-skill/. For Claude Code, copy it into .claude/skills/scrum-skill/ and invoke /scrum-skill. Preserve its scripts, references and assets. Do not overwrite an existing installation without reviewing the change.
Skills-only installation does not register a plugin, hooks or an agent runtime. See Codex skill discovery. This repo publishes compatible metadata; it does not install anything into your account automatically.
Use explicit modes or a role-specific request; a bare call opens a guided wizard.
$scrum-skill wizard
$scrum-skill setup --goal "Staff can inspect stock changes" --roles engineering,qa,design
$scrum-skill resume
$scrum-skill plan
$scrum-skill deliver --target US-7
$scrum-skill role design plan --target apps/web
$scrum-skill role legal-counsel review --target docs/architecture
$scrum-skill role qa review --target .github/workflows/ci.yml
Invocation guide explains modes, role aliases, the guided wizard and the optional intent resolver. These argument routes use one canonical skill; they do not register separate workers or automatically run tools. Natural-language requests continue to work:
$scrum-skill Resume from the project's latest handoff and reconcile the actual state.
$scrum-skill Refine these stories, identify evidence gaps and plan the next Sprint.
$scrum-skill Plan and start the next authorized slice; coordinate specialists and verify it.
$scrum-skill Prepare a Sprint Review using actual results and separate unfinished work.
$scrum-skill Run a retrospective and propose a few measurable improvements.
Use /scrum-skill:scrum-skill for the plugin version in Claude Code. Setup/planning creates artifacts; it does not by itself authorize implementation, pushes, deployment or messages. If your host cannot delegate, the coordinator works sequentially and records that mode.
The plugin includes an offline dashboard skill. Direct folder installations must also
copy skills/scrum-dashboard/ into the host's skill directory to use it.
$scrum-skill dashboard
$scrum-dashboard Generate a local dashboard from .scrum/metrics.json.
Claude plugin: /scrum-skill:scrum-dashboard. The dashboard shows project burndown
by Sprint, completed Sprint velocity, changing backlog scope, progress and a dated
Gantt timeline. It has no daily chart or separate burnup. Missing dates, estimates and
history remain visible; agent speed does not invent a two-week calendar.
Dashboard schema and commands explain
how to supply snapshots. New workspaces start with an empty metrics.json; the helper
does not infer history from Markdown. Generate a labeled synthetic demo with:
python skills/scrum-dashboard/scripts/generate_dashboard.py --demo --output demo.htmlPreview the synthetic dashboard.
Open the resulting HTML locally. It needs no server or external libraries and embeds all input data, so sharing it shares that data.
Stories now have explicit point/estimate-state fields and an estimate decision ledger. Estimation and Planning Poker cover actual participants, private votes, discussion and agreement, provisional agent estimates, capacity, unfinished work and frozen historical velocity. Points and Planning Poker are supporting practices, not mandatory Scrum rules.
Project lifecycle guide connects discovery, refinement, delivery, reviews, releases and operational handover. Workflow review records coverage and remaining limits.
Scrum uses transparency, inspection and adaptation to work toward a Product Goal. These diagrams summarize the framework; source attribution and packaging references distinguish Scrum rules from this package's AI/project conventions. Using only selected practices is Scrum-informed, not a claim of full Scrum implementation or certification.
flowchart TB
Goal["Product Goal"] --> PB["Ordered Product Backlog"]
PB --> Plan["Sprint Planning: why, what, how"]
Plan --> SB["Sprint Goal and Sprint Backlog"]
SB --> Work["Developers build, verify and adapt"]
Daily["Daily Scrum: inspect Goal progress"] --> Work
Work --> Quality{"Meets shared Definition of Done?"}
Quality -->|Yes| Inc["Usable integrated Increment"]
Quality -->|No| Active{"Sprint still active?"}
Active -->|Yes: adapt and continue| Work
Active -->|No: unfinished at Sprint end| PB
Inc --> Review["Sprint Review: inspect product and adapt backlog"]
Review --> PB
Review --> Retro["Retrospective: improve quality and effectiveness"]
Retro --> Plan
Inc -.->|When release policy permits| Delivery["Delivery; Review is not a release gate"]
One active Sprint, fixed length of one month or less. The team chooses a useful cadence; two weeks is a starting convention. The Developers own their plan and estimates. A timebox ending does not turn unfinished work into a Done Increment.
flowchart LR
subgraph Team["One Scrum Team — no sub-teams or internal hierarchy"]
direction LR
PO["One Product Owner<br/>Value, Product Goal and backlog order"]
Dev["Developers<br/>Decide who, when and how<br/>Engineering, QA, design and other skills"]
SM["One Scrum Master<br/>Scrum practice and team effectiveness"]
PO <-->|"Goal, priorities and tradeoffs / feasibility and feedback"| Dev
SM ---|"Coach self-management; help remove impediments"| Dev
SM ---|"Support goals, backlog and stakeholder collaboration"| PO
end
The Product Owner is one person; the Scrum Master serves the whole team. Developers own their plan, estimates and implementation choices. Layout and connecting lines express collaboration, not reporting authority. AI workers assist these accountabilities; a generated profile is not a human appointment, stakeholder participation or regulated professional sign-off. This package deliberately keeps a human accountable for product decisions.
Legal Counsel reviews platform/design choices, flags legal risk and advises on defensible options and safeguards. Privacy counsel focuses on applicable data-handling requirements. Their output guides product/design decisions; source-backed human judgment remains necessary.
Identify the Product Owner and Scrum Master from actual instructions or project records. Skill invocation does not appoint anyone. If either is unknown, record it as unresolved and call the current setup Scrum-informed. The optional helper accepts --product-owner and --scrum-master for identities already supplied; neither proves appointment, participation or full Scrum implementation.
Read accountability guides for Product Owner, Scrum Master and Developers behavior and decision boundaries. Role selection links presets for engineering, QA, UI/UX, security, optional Legal Counsel and PDPA/GDPR Counsel, AI/data, operations and domain review. Each preset describes its contribution, working approach, evidence and boundaries. Specialisms share one Definition of Done and collaborate throughout delivery.
Setup copies selected presets into .scrum/roles/; unknown role names use the adaptable template. Existing profiles are preserved. These are portable instructions, not automatically registered agents. For example: --specialists engineering,qa,design,security; ask a delegated worker to read its profile, the shared DoD and its bounded task when delegation is authorized.
On a clean run, the first-run setup identifies the selected team's required imports, presents concrete commands and scope, and verifies availability after installation. Legal roles select LQ bundles; web design selects taste/minimalist/redesign, Impeccable and applicable GSAP/Lenis libraries. Full LQ catalog import is an explicit option. Required missing tools keep the affected workflow unready until imported or an explicit fallback is chosen.
scrum-skill contains original instructions, profiles and a workspace helper. It does not vendor LQ, taste-skill, Impeccable, GSAP or Lenis source, install those dependencies, or register specialist agents. The routing protocol requires resolving actual host availability, loading the selected skill and recording what ran or was unavailable. Source links alone are not executable integration. GSAP and Lenis are application libraries, not agent skills.
flowchart TB
PB["Product Backlog: ordered work and refinement"] --> PG["Commitment: Product Goal"]
SB["Sprint Backlog: selected work and adaptable plan"] --> SG["Commitment: Sprint Goal"]
INC["Increment: usable integrated outcome"] --> DOD["Commitment: Definition of Done"]
AC["Supporting practice: item acceptance criteria<br/>Not a formal Scrum artifact"] -.->|Behavior to verify| INC
Evidence["Supporting practice: revision-specific evidence<br/>Not a formal Scrum artifact"] -.->|Quality to demonstrate| DOD
Acceptance criteria describe each item's behavior. The Definition of Done describes shared product quality. Acceptance criteria and evidence maps are supporting practices, not additional formal Scrum artifacts. CI, architecture, accessibility, security/privacy and domain review are chosen for the actual project. Local tests, mocks and coverage do not establish every integration or human decision.
| Event | Inspect/adapt | Record actual participation or label preparation |
|---|---|---|
| Sprint Planning | Why this Sprint matters, what is feasible, how to deliver | Goal, selected items, capacity/uncertainty and plan |
| Daily Scrum | Progress toward the Goal and the Developers' next plan | Evidence, plan changes and impediments |
| Sprint Review | Product outcome, real stakeholder feedback and backlog order | Done results; unfinished work listed separately |
| Retrospective | Collaboration, tools, quality and assumptions | Owned improvements with observable success measures |
Review is not merely a demo; retrospective is not product acceptance. Async agent notes help prepare events but do not prove a human ceremony occurred. Never fabricate attendance, velocity, passed tests or approval.
flowchart TB
Invoke["Invoke skill with project outcome"] --> Inspect["Inspect instructions, repo and existing state"]
Inspect --> Existing{"Existing Scrum records?"}
Existing -->|Yes| Resume["Resume and reconcile; preserve decisions"]
Existing -->|No| Init["Create adaptable local workspace"]
Resume --> Refine["Refine Goal, team, quality and first slice"]
Init --> Refine
Refine --> Scope{"Delivery requested and authorized?"}
Scope -->|No| Artifacts["Reviewable plan and open questions"]
Scope -->|Yes| Deliver["Bounded build, verification and integration"]
Artifacts --> Handoff["Evidence-linked handoff and next action"]
Deliver --> Handoff
The optional helper creates these records, all initially proposed/preparatory:
project/
AGENTS.md Preserved; a small marked resume pointer is appended
CLAUDE.md Preserved; the same pointer is appended
.gitignore Preserved; .scrum/ ignored by default
.scrum/
workspace.json Skill version and project configuration
TEAM.md, roles/ Adaptable specialist profiles
ACCOUNTABILITIES.md Product Owner, Scrum Master and Developers behavior
PRODUCT_GOAL.md, PRODUCT_BACKLOG.md, DEFINITION_OF_DONE.md
DECISIONS.md, IMPEDIMENTS.md, EVIDENCE.md
sprints/ Forecast, planning, review and retrospective
handoffs/INITIAL.md, handoffs/LATEST.md
MASTER_PROMPT.md Fresh-session resume instructions
For a manual setup preview from a clone:
python skills/scrum-skill/scripts/init_workspace.py --root /path/to/project --name inventory --goal "Staff can track stock changes" --specialists engineering,qa,design --dry-runRun without --dry-run only when you intend setup. Use --specialists none for a sequential generalist. Existing records are not overwritten; malformed pointer blocks, conflicting project identity and path escapes fail before writes. Choose an explicit --start-date YYYY-MM-DD in the project's timezone when dates matter. The helper does not run tests, create a Git repo, push, spawn agents or deploy; the skill guides authorized work afterward.
flowchart TB
Observe["Observed project friction"] --> Sanitize["Generalize; keep private facts local"]
Sanitize --> Repro["Invented minimal reproduction"]
Repro --> Change["Narrow skill or template change"]
Change --> Review["Scrum Master or QA review"]
Review --> Check["Behavior and package checks"]
Check --> Version["Versioned release"]
Version --> Adopt["Consumer reviews migration; no silent sync"]
Keep one canonical skill under skills/scrum-skill. Both host manifests and catalogs point to it; there are no generated copies to drift. A project's domain safeguard or tool choice becomes an optional profile, not a universal rule. See maintenance guidance and contributing.
python scripts/validate_package.py
python -m unittest discover -s tests -v
claude plugin validate .claude-plugin/plugin.json
claude plugin validate .claude-plugin/marketplace.jsonPackage tests check manifest paths/versions, references and helper behavior. They do not certify facilitation quality or claim a live installed session on every host. The evaluation record describes what was actually exercised. Mermaid README sources are rendered during authoring to catch syntax/layout failures. No supplied PDFs or project-specific backlog content is part of this repo.
Created by Kevan Wee and contributors, licensed under MIT. See contributing and help-wanted issues.