fredantb/spec-driven-development

This is a Claude skill called Spec-Driven Development that addresses a common problem when using multiple AI coding tools together. When you ask Claude Code to build something, then ask Cursor to fix it, then ask Copilot to clean it up, each tool makes its own assumptions about what you want because they have no shared reference point. This skill creates that shared reference point before any code is written. The skill works by interviewing you with a few short questions, then generating three planning documents: requirements.md (what the system must do, using traceable REQ-xxx identifiers), design.md (how the system will be built), and tasks.md (an ordered list of atomic implementation steps, each linked back to a specific requirement). Every AI tool you use then reads these files before doing anything, which stops them from inventing conflicting interpretations. Once the spec files are in place, the skill also generates matching configuration files for each AI tool you use, including Claude Code's CLAUDE.md, Cursor's .cursorrules, Windsurf's .windsurfrules, GitHub Copilot's instructions file, and Aider's config. All of these files contain an identical instruction block that tells the AI to read the spec files first, never implement requirements that are not written down, and stop and ask before deviating from the design. The skill supports two starting points: new projects (where it interviews you and builds the spec from scratch) and existing codebases (where it reverse-engineers a spec from what you describe, marking inferred fields as needing verification). It ships with a test suite of 135 assertions and a GitHub Actions workflow that checks spec quality on every push. The project is licensed under MIT and is currently in beta.