EvanFlow provides a structured, TDD-driven iterative feedback loop for AI-assisted software development within Claude Code. It addresses common AI coding failures by enforcing checkpoints, rigorous testing, and user control, benefiting engineers seeking quality and predictability in AI-generated code.

How It Works

The system orchestrates ideas through brainstorm, plan, execute, and iterate phases using 16 Claude Code skills. It integrates vertical-slice TDD (RED→GREEN→REFACTOR) within code tasks and requires explicit user approval at critical checkpoints. For independent units, it employs parallel orchestration with coder/overseer subagents and integration tests to maintain contracts.

Quick Start & Requirements

Install via Claude Code plugin marketplace: /plugin marketplace add evanklem/evanflow, /plugin install evanflow@evanflow, then restart. Requires Claude Code, Bash, and jq. chromium is optional for UI visual verification.

Highlighted Details

• Git Guardrails: Auto-activated block-dangerous-git.sh hook prevents destructive Git operations via Bash tool.
• Assertion Correctness: Explicit checks validate LLM-generated test assertions to prevent common failures.
• Five Failure Modes: Iteration phase checks for hallucinated actions, scope creep, cascading errors, context loss, and tool misuse.
• Parallel Orchestration: Forks into parallel coder/overseer subagents for independent units, enforcing contracts via integration tests.

Maintenance & Community

Contributions reducing ceremony or adding evidence-backed improvements are welcomed, aligning with the project's opinionated design. Specific community channels are not detailed.

Licensing & Compatibility

Licensed under MIT, permitting broad commercial use and linking within closed-source projects.

Limitations & Caveats

Highly opinionated against added ceremony or auto-actions, prioritizing user control. Requires explicit user direction at checkpoints. Dependency failures (e.g., jq) can compromise safety features like the Git guardrail.