As GitHub Copilot agents become part of everyday development, a new bottleneck has emerged: unclear intent. When Copilot is driven by vague or non‑deterministic prompts, it generates code that is inconsistent, brittle, and misaligned with project architecture. This problem is amplified in community‑driven FOSS projects, where architectural rules are often implicit or undocumented.

​In this session, we share lessons from building and maintaining Visage—our primary Spec Kit–driven platform—along with hackmum.in, our community website, by adopting Spec‑Driven Development (SDD) as the control layer for Copilot agents. Using Spec Kit and the lightweight OpenSpec framework, we show how specifications become the primary instruction set for AI—before Copilot writes a single line of code.

​Instead of prompt‑first development, we demonstrate spec‑first, agent‑executed workflows that turn Copilot from a guessing engine into a predictable collaborator.

​What you’ll see

​Takeaways

​Attendees will leave with an open‑source blueprint for using GitHub Copilot agents responsibly:

​Improve determinism in AI‑assisted coding

​Preserve architectural intent in community projects

​Scale contributions without increasing review burden

​This talk is for developers who want Copilot agents to amplify engineering discipline - not replace it.