A Claude Code plugin named team, with skills, subagents, slash commands, scripts, and templates. The subagents (engineer, reviewer, debugger, and more) combine into workflow recipes.
This repo is a Claude Code plugin with agent definitions, slash commands, and workflow skills. It is also its own marketplace, named andymai. When you install the plugin, Claude Code registers its agents in every project that you open.
You need Claude Code. In any session, run:
/plugin marketplace add andymai/skills
/plugin install team@andymai
To get later changes:
- Refresh the catalog with
/plugin marketplace update andymai. - From a shell, run
claude plugin update team@andymai, or pick Update from the/pluginmenu.
To remove everything, run /plugin uninstall team.
The plugin name is the namespace of each component. Commands and skills are /team:commit, /team:shepherd-pr, and so on. Agents are team:engineer, team:reviewer, and so on.
To work on the plugin itself, load a clone for one session with claude --plugin-dir /path/to/skills. An install copies the repo into the plugin cache, so edits to a clone do not reach the installed copy.
| Agent | Purpose | Model | Features |
|---|---|---|---|
| engineer | Implement features and their tests to match the existing code patterns. Write the failing test first for a bug fix | Opus | |
| debugger | Investigate a bug: reproduce it, fix all layers, and use only real dependencies | Opus | |
| reviewer | Review code, and report only findings with a confidence of 60 or more | Opus | read-only |
| researcher | Explore codebases, compare technologies, gather information | Opus | read-only |
| gap-finder | Verify that implementations match specs, and find missing requirements | Opus | read-only |
| optimizer | Simplify and refactor recently changed code, with no change in behavior | Sonnet |
For planning, use plan mode or Claude Code's built-in Plan agent. For a security review, use the built-in /security-review.
By design, this plugin does not:
- Replace or install Claude Code. It needs Claude Code installed and logged in. This plugin only adds agents, slash commands, and skills on top of it.
- Provide a runtime orchestration framework. The agents are static markdown definitions that Claude Code reads. There is no daemon, scheduler, or message bus between agents.
- Make workflow recipes executable. The recipes under Workflow Recipes suggest an order in which to invoke the agents. They are not automated pipelines that you can run with one command.
- Bootstrap new projects. The plugin adds agents and commands for existing projects. It does not scaffold repos, generate boilerplate code, or configure CI. The obsidian-project skill writes planning notes in an Obsidian vault, not code.
- Work outside Claude Code. The agent and slash-command formats are specific to the subagent protocol of Claude Code. Other LLM tools cannot use them without changes.
Bug fix:
debugger → engineer → reviewer
engineer → gap-finder → reviewer
Cleanup after a change:
optimizer → reviewer
Research before a decision:
researcher (standalone)
Each agent works on its own and returns results. Claude Code decides which agent to invoke next.
These Agent Skills are not specific to one stack. After you install the plugin, every project can use them. Like all skills, they load in steps: only the description stays in context until the skill triggers.
| Skill | Triggers on | What it does |
|---|---|---|
| shepherd-pr | "shepherd the PR", "drive this PR to green", "clean up the PR" | Runs an autonomous loop on the PR of the current branch. It resolves review comments, fixes check-run findings (including neutral-status reviewers that the CI rollup hides), and waits for CI to pass. Then it reassigns the PR and reports. It works in any repo: it detects Graphite or plain git and gets the GitHub login at runtime |
| outward-writing | the deliverable is prose another person will read: a design doc, proposal, wiki page, Slack message, PR body or reply, review reply, ticket, email | You collect every fact with a source, write the draft in ASD-STE100 Simplified Technical English, fact-check it sentence by sentence, and run check.py. You publish nothing until the user sees the draft. It needs no external CLI |
| obsidian-project | "start a project in my vault", "set up project docs in Obsidian" | Creates one folder in an Obsidian vault with five linked notes: the overview, PRD, architecture, research, and tasks. It adds the project to the vault's index. It fills a section only from what you said or what it read, and the other sections keep a hidden prompt. A layout in the vault's own CLAUDE.md wins over the templates |
| interview | "grill me", "interview me", "stress-test this plan" | Interviews you in rounds until each open decision is settled. It looks up the facts itself and asks you only for decisions, each with a recommended answer. It restates the result and stops when you confirm it. It writes no plan and no code |
| agent-docs | you create, change, or review a skill, a subagent, a command, or a CLAUDE.md | Gives the authoring rules for a document that an agent follows: write for the weakest executor, write the description as a trigger, keep one source for each fact. Its reference file has the Claude Code frontmatter keys, argument tokens, and limits |
| handoff | you type /team:handoff |
Writes a handoff document outside the repo, so that a new session can continue the work. It labels each claim as verified, inferred, or assumed, and it gives one exact next step |
| retrospective | you type /team:retrospective |
Reviews a session and proposes changes to the agent's environment. Each proposal quotes the moment in the session that shows the cost. A mechanical rule becomes an automated check, not a line in a rules file |
| tickets | you type /team:tickets |
Splits a plan or a spec into tickets. Each ticket is a vertical slice with acceptance criteria and the tickets that block it. It writes local files, or it publishes to the tracker that you or the project name |
| Command | Description |
|---|---|
/team:audit-claudemd |
Audit or bootstrap the project's CLAUDE.md against observed conventions |
/team:branch |
Create a branch with <type>/<kebab-description> naming |
/team:check |
Run the project's local quality gate (auto-detected) |
/team:commit |
Create a conventional commit from working tree changes |
/team:contribution-report |
GitHub contribution summaries for performance reviews |
/team:pr-description |
Generate PR title and description from branch changes |
/team:upgrade-dep |
Upgrade a dependency and fix breaking changes |
/team:worktree |
Set up a git worktree under .worktrees/ for parallel work |
Claude can also start a command on its own when your request matches the command's description. The exception is /team:contribution-report, which runs only when you type it, because it creates a GitHub repo and uploads to it.
skills/outward-writing/scripts/check.py ships with the outward-writing skill, which reaches it through ${CLAUDE_PLUGIN_ROOT}. It fails a draft on em or en dashes outside code and quotes. It also fails a sentence over 25 words and a paragraph over 6 sentences. It warns on the present perfect, "-ing" verbs, possible passives, and two verbs for one action, and it reports the word count.
Five skills adapt ideas from two MIT-licensed projects. The text in this repo is new. THIRD-PARTY-NOTICES.md has the license notices.
- interview: the decision rounds come from
grillingin mattpocock/skills. The restatement and the explicit confirmation come frominterview-mein addyosmani/agent-skills. - agent-docs, handoff, retrospective, tickets: from
writing-for-agents,handoff,retro, andto-ticketsin mattpocock/skills.
For Obsidian syntax, Bases, Canvas, and the Obsidian CLI, install the official obsidianmd/obsidian-skills plugin. The obsidian-project skill uses its skills when they are present.
To add or upgrade an agent, read AGENTS.md, the playbook for agent authors. It covers the frontmatter schema, model selection, and the color palette. It also states two rules:
- The generalization rule: agents run in every repo you touch, so each rule must be portable.
- The weakest-executor rule: prompts must carry the discipline (evidence rules, hard gates, self-checks), so smaller models and junior engineers also succeed with them.
Run these gates before you open a PR. CI runs the same four:
python3 -I scripts/lint.py --base origin/main
python3 -I scripts/tests/test_lint.py
skills/outward-writing/tests/run.sh
claude plugin validate . --strict
The lint checks the frontmatter, the names, the cross-references, and the README tables. It also fails when a shipped file changes and the version in .claude-plugin/plugin.json stays the same.
The cases under evals/ test that each skill starts on a request that it owns and stays quiet on a nearby request. Each run is a real model call on your account, so CI does not run them. Run them after you change a skill's description:
claude plugin eval . --model sonnet --no-publish