Summary

CCteam-creator orchestrates multiple AI agents within Claude Code, functioning as a coordinated engineering team. It targets developers and power users seeking to scale AI-assisted development by enabling parallel agent teams with robust file-based planning, role-based collaboration, and integrated CI/CD-like enforcement.

How It Works

A "Team-Lead" agent manages specialized AI agents using file-based planning (task_plan.md, findings.md, progress.md) for persistent state survival. Agents communicate via direct messages and file persistence, employing protocols like 3-Strike Escalation and Guardrail Capture. Key innovations include Doc-Code Sync, Invariant-Driven Review, and a "Taste Feedback Loop" translating user preferences into automated enforcement, drawing from OpenAI's harness engineering and Anthropic's multi-agent research.

Quick Start & Requirements

• Prerequisites: Claude Code experimental agent teams must be enabled (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1).
• Installation: Via Claude Code Marketplace or manual Git clone/copy to ~/.claude/skills/.
• Usage: Invoke with /CCteam-creator or natural language prompts.
• Links: GitHub Repository

Highlighted Details

• Team-Lead Control Plane: Manages user alignment, phase gates, and project files.
• Doc-Code Sync: Enforces documentation updates alongside code changes, verified by the reviewer.
• Invariant-Driven Review: Code checked against formal system invariants, promoting automated tests.
• Failure-to-Guardrail Loop: Resolved failures systematically captured as permanent automated checks.
• Taste Feedback Loop: User style preferences encoded into automated enforcement.
• Team Snapshot: Enables near-instantaneous project resumption via cached agent onboarding.
• Built-in Agent Protocols: Suite of predefined communication and workflow protocols.

Maintenance & Community

No specific community channels or detailed contributor information provided in the README.

Licensing & Compatibility

• License: MIT License.
• Compatibility: Permissive, suitable for commercial use.

Limitations & Caveats

Claude Code's context compaction can degrade performance with large contexts (1M); default 200k is recommended. A workaround involves exiting/resuming and reading team-snapshot.md if compaction occurs, as the team-lead may temporarily lose operational memory. Marketplace plugins require manual updates, though an in-app notification system exists.