DevSpark: Constitution-Driven AI for Software Development

The DevSpark series documents how I built and use a constitution-driven toolkit for AI-assisted software development. Rather than relying on generic AI suggestions, DevSpark aligns coding agents — GitHub Copilot, Claude Code, Cursor, and others — with a project's specific architectural principles, quality standards, and decision history.

The series tells a story in roughly chronological order: why the problem of ungoverned AI assistance is real, how the initial specification-driven workflow evolved into a full governance framework, what building it in practice actually looked like, and where it has grown since. Each article stands on its own, but read in sequence they trace the evolution from a simple fork to a framework now used across multiple production codebases.

The Core Idea

Every software project has rules. Most are implicit — living in individual developers' heads, scattered across old PR comments, encoded in conventions nobody has written down. When AI coding agents generate code without access to those rules, they produce output that is technically correct but architecturally wrong. The fix isn't a better prompt. It's a constitution: a formal, version-controlled document that captures the non-negotiable principles, encodes the decisions already made, and gives every AI agent the context it needs to generate code that actually fits the project.

DevSpark makes that constitution the center of gravity for every part of the development workflow: specifications, plans, implementation, PR reviews, site audits, and now multi-step harness workflows. The constitutional principles aren't aspirational guidelines — they're enforced mechanically at every stage.

The DevSpark Repository

The full toolkit — prompts, workflow specs, scripts, and documentation — is available at github.com/markhazleton/devspark. DevSpark currently supports 17 AI coding agents through a centralized registry, with shims auto-generated for each platform from canonical prompt templates.

Contributing

If you encounter a prompt that consistently produces wrong output in a reproducible way, the right place for that is the DevSpark issue tracker. Include the command name, your project context, what the agent produced, and what you expected. Specific, reproducible cases are what drive the framework forward.