ww-w-ai/bkit-claude-code

What it solves

bkit is a plugin for Claude Code that transforms it into a structured development system. It solves the problem of "vibe-coding"—where AI generates plausible-looking code that may not actually match the original design or requirements—by introducing rigorous verification, context management, and a multi-agent orchestration workflow.

How it works

bkit implements a "Context Engineering" approach using a set of specialized agents and quality gates. It operates through three primary commands:

• /sprint: Manages multi-feature releases by splitting them into "context-budgeted" sprints (typically $\le$ 75K tokens) to prevent the AI from losing focus or running out of memory. It uses a master plan to track progress across sessions.
• /pdca: Executes a Plan-Do-Check-Act loop for single features. This involves a team of specialist agents (PM, CTO, Developer, QA) who create PRDs, design docs, and code, then verify the result against the design.
• /control: A single autonomy dial (Levels 0-4) that determines how much of the process runs unattended, from fully manual to fully autonomous.

To ensure quality, bkit uses Quality Gates (like the M1 matchRate) that measure the alignment between the design spec and the generated code. If the match rate is below 90%, the system automatically triggers a self-repair cycle.

Who it’s for

• First-time AI coders: Those who want to ship production-quality software without needing deep senior engineering intuition.
• Solo developers and indie builders: Those who need a "team-in-a-box" to handle frontend, backend, QA, and security in parallel.
• Team leads: Those managing complex releases that exceed a single AI session's context window.
• Non-English speakers: Users who can describe requirements in multiple languages (8-language auto-detection).

Highlights

• Context-Budgeted Sprints: Prevents context window overflow by splitting projects into manageable, dependency-aware chunks.
• Coded Documentation: Follows a "Docs = Code" philosophy where every feature produces a PRD, design, and report, which then serve as the contract for verification.
• Automated Self-Repair: Uses a gap-detector and pdca-iterator to automatically fix code that drifts from the design spec.
• Multi-Agent Orchestration: Spawns specialist agents (e.g., pm-lead, cto-lead, qa-lead) to handle different phases of the development lifecycle.
• Detailed Quality Gates: Includes 11 specific gates, including a 7-layer dataFlow integrity check.