nnguyenquangg/claude-code-spec-driven-development

This is a toolkit that adds a structured development workflow to Claude Code, an AI coding agent. The core idea is spec-driven development: before any code gets written, the system first produces a written specification and a set of architecture decision records, then pauses for a human to review and approve them. Only after approval does it move on to building. This two-phase split is the main thing that distinguishes it from the default Claude Code experience, where planning and building blur together. The toolkit installs five commands. The /make-plan command takes a description of a feature or change and works through a clarification step (asking questions if the request is vague), produces a proposal and design, writes down the significant decisions it is making, and then stops. You review what it produced. Once you are satisfied, /implement-specs picks up the approved specs and runs a loop: implement, review independently, fix problems, repeat, up to six cycles, then run linting, tests, and a build check. The /fix command handles bug reports, diagnosing the root cause and applying a minimal correction. The /research command investigates an unfamiliar idea by searching the web and the codebase, producing a brief with options. The /autopilot command chains both phases together for situations where you want less back-and-forth, replacing the human review gate with an AI review of the specs. Under the hood, each command routes to specialist skills. Depending on the task, it may invoke a clarification skill, a proposal generator, stack-specific experts for NestJS, PostgreSQL, React, or TypeScript, a code verifier, and a quality fixer. You do not have to pick which experts to use: the system reads the task and selects them. Installation runs a shell script that symlinks the skills and commands into your Claude Code configuration folder. It also installs four external dependencies: the OpenSpec CLI, a clarification skill from a separate project, and two plugin packs that supply the language-specific experts and code verification tools. The project is MIT licensed and covers only the skills and commands in this repository. The external dependencies each carry their own licenses.