Skip to content

Latest commit

 

History

13 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

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.

Author

Luke Steuber · luke@lukesteuber.com · https://lukesteuber.com

About

Council-style codebase-to-pitch plugin for Claude Code: 12 seats across four subcommittees decide what your code is, who buys it, and what kills it.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages