Team

Council-style codebase-assessment plugin for Claude Code, Codex, and Cursor.

Point it at a codebase. An 11-seat team across three subcommittees (research, technical, skeptics) plus a standalone legal red-flag check decides what it is, where it fits, what breaks it, what kills it, and how to know whether the assessment is any good.

The package contains 6 skills and 11 agent definitions. The manifests for all three supported runtimes point at the same skills/ directory, so the skill inventory has one source of truth.

Install

Claude Code

/plugin marketplace add actually-useful-ai/team
/plugin install team@lukeslp-team

Codex

Import https://github.com/actually-useful-ai/team from the Plugin Directory, then install Team from the imported marketplace.

Cursor

cursor-agent plugin marketplace add https://github.com/actually-useful-ai/team
cursor-agent

In the interactive agent, open /plugin, install Team at user scope, and it will be available to both the IDE and CLI.

Each runtime has its own manifest; all three load the shared skills/ directory. Cursor also loads the shared agents/ directory explicitly.

Usage

/team                            # assess the current directory
/team src/                       # assess a specific path
/team ~/projects/whatcolor       # assess a project elsewhere on disk
/team "Should I open-source this?"   # team debate on a question

What you get back

A short verdict from the executive chair, structured as:

• Decision: one line.
• The assessment: what it is, where it fits, what it's good for.
• What kills it: kill criteria, top three risks, the strongest attack.
• Verification plan: how to know if the assessment is right.
• Legal red flags: license/IP blockers, or all-clear.
• Dissenting opinions: preserved, not flattened.
• Consultant notes: what external models said when they could be reached.

The seats

Research (facts, not opinions)

• recon: maps the codebase. Languages, dependencies, public surfaces, internal patterns.
• scout: external prior art. Comparable projects, technical approaches, and verifiable current sources; optional Craft Ask leads when explicitly authorized.

Technical

• architect: fit with the existing codebase patterns. Chairs the committee.
• greybeard: old-engineer review of what breaks at 10x. Encodes the Google SRE Book (chapter 22, four golden signals), Brendan Gregg's USE method. Optional consult: Codex, Grok.
• safety: failure recovery, fallbacks, graceful degradation.
• tester: verification strategy, regression criteria.

Skeptics

• breaker: adversarial attack on the assessment. Concrete failure scenarios only.
• cynic: pre-mortem (Klein 2007), kill criteria (Annie Duke), Amazon working-backwards critique. Must always disagree. Optional consult: full /consensus fan-out.

Legal check (standalone)

• legal: IP, licensing, attribution. A red-flag list of license/IP blockers that would stop the project from being shipped or open-sourced. Looks at every dep's license and the project's own.

Chair and polish

• executive: reads the committee reports plus the legal list, synthesizes the verdict, preserves dissent.
• editor: strips machine-generated writing indicators. Post-verdict only.

Consultants

Several seats can use native read-only agents for diverse priors. When Luke explicitly authorizes outside-model consultation or fan-out, Team can also use Craft's optional craft-ask entry point. Craft owns provider/model routing and provenance; Team retains its native fallback and works without Craft or external credentials.

The full fan-out lives at skills/consensus/ and is invoked by the cynic seat as the panel of last resort.

Why a council and not a single pass

A single LLM pass will produce an assessment that sounds confident regardless of merit. Adversarial review: especially from seats anchored to canonical frameworks (SRE Book, USE method, Klein pre-mortem, Annie Duke kill criteria), catches the failure modes a confident single pass misses. The subcommittee structure prevents the executive from drowning in eleven separate voices.

The design is sympathetic to the parallel-critics + single-synthesis pattern that holds up better in recent multi-agent research (arxiv 2509.05396, 2025) than round-table debate.

Native CLI consensus

Consensus selects Claude Code, Grok, Z.ai through an isolated Claude Code configuration, Muse Code, and Ollama Cloud through its native CLI. When Craft is installed, craft-ask --list owns route discovery. A missing CLI or login remains unavailable; no API or gateway substitution is made. Requested model names remain distinct from CLI-reported identity, and aliases of one provider do not count as independent voices.

License

MIT.

Around here

Luke Steuber · Data Poems · Ambient Time · Actually Useful AI · One Impossible Thing

Made by Luke Steuber. Questions or collaboration: luke@lukesteuber.com.