Brewcode 3.4: Teams, Debate, and Design-to-Code
Update, October 2026: this post describes an earlier Brewcode version, and some of the features and commands it presents are retired. See the current docs.
Brewcode 3.4 ships three features that change how you work with Claude Code at scale: dynamic teams, multi-agent debate, and design-to-code conversion. Each one addresses a real bottleneck in AI-assisted development workflows.
Here is what each feature does and why it matters.
Dynamic Team Creation
Teams let you define a group of specialized agents that collaborate on a shared codebase. In 3.4, teams are dynamic — agents self-select tasks based on their domain expertise, and all activity is traced to a JSONL file for auditability.
Define a team in your project configuration:
team: default
agents:
- name: site-dev
domain: Astro/Tailwind/DaisyUI
mission: Build pages, components, layouts
- name: devops
domain: Docker/Caddy/CI
mission: Containers, workflows, deploy
- name: docs-writer
domain: Operational docs
mission: Runbooks, onboarding, server docs
Start the team with a single command:
/brewcode:teams start
Each agent picks up tasks that match its domain. The manager agent coordinates, resolves conflicts, and ensures no two agents edit the same file simultaneously. Every action is logged to .claude/teams/default/trace.jsonl so you can replay, audit, or debug any decision.
The key insight is that you describe what each agent is good at, not what it should do. Task assignment happens dynamically based on the work queue.
Multi-Agent Debate
Code review by a single AI reviewer tends to produce either too many false positives or miss real issues. Debate changes this by running multiple reviewer agents in parallel, each with a different perspective.
The debate flow works in three phases. First, each reviewer evaluates the code independently. Second, reviewers share their findings and challenge each other’s conclusions. Third, a quorum vote determines which issues are real.
Configure the quorum threshold:
/brewcode:review --debate --quorum 2/3
With a 2/3 quorum, an issue is flagged only if at least two out of three reviewers agree it is a genuine problem. This eliminates the noise of single-reviewer false positives while catching issues that one reviewer might miss.
In our testing across 200+ pull requests, debate-mode review caught 40% more real bugs than single-reviewer mode while producing 60% fewer false positives. The tradeoff is higher token usage — roughly 3x a single review — but the quality improvement justifies the cost for production code.
Design-to-Code Conversion
The new /brewcode:design skill uses a GLM vision model to convert screenshots and mockups into working components. Point it at an image file and specify the target framework:
/brewcode:design screenshot.png --framework astro --css tailwind
The vision model analyzes layout, spacing, typography, and color, then generates a component that matches the design. It respects your project’s existing design tokens — if you have CSS custom properties defined, the generated code uses them instead of hardcoded values.
This is not pixel-perfect reproduction. It is a strong starting point that gets you 80-90% of the way there, which you then refine manually. The alternative is writing every component from scratch while constantly switching between your editor and the design file.
Design-to-code works best with clean, well-structured mockups. Busy screenshots with overlapping elements or inconsistent spacing produce weaker results. When possible, export individual components or sections rather than full-page screenshots.
Upgrading to 3.4
Update Brewcode from the Claude Code marketplace:
claude plugin marketplace update claude-brewcode
claude plugin update brewcode@claude-brewcode
Teams require at least Claude Code 2.1 with background agent support. Debate mode works on any Claude Code version. Design-to-code requires a vision-capable model in your configuration.
All three features are documented in detail at doc-claude.brewcode.app. If you run into issues, open a discussion on the GitHub repo.