LLM Development
Proven practices for building software with AI assistance
This site distills insights from analyzing real-world projects built with GitHub Copilot and other AI coding assistants. Learn what actually works—backed by evidence from production codebases.
Quick Navigation
Evidence-Based Insights
Every practice on this site is backed by analysis of real repositories:
- 4 comprehensive case studies (180+ PRs, 900+ commits analyzed)
- Proven velocity patterns: 3-11 PRs/day sustained for weeks
- Measurable outcomes: 1-3 commits for 80% of well-scoped work
- Production results: From prototype to production in weeks
What Makes This Different
Most AI coding advice is speculation. This site is built on systematic analysis of what actually works in production:
✅ Specific examples from real PRs
✅ Metrics that matter (velocity, iteration, quality)
✅ Anti-patterns to avoid
✅ Actionable templates you can use today
About This Site
This site was built using its own methodology:
- Clear Vision: Started with VISION.md defining the mission and goals
- Design Analysis: Evaluated multiple approaches before implementation
- Tiny MVP: Built a minimal working kernel (phased instruction approach)
- Iterative Refinement: Enhanced through real-world use and feedback
The content comes from the llmdev project—tools for analyzing LLM-assisted development patterns.
Learn more about how this site was built →
Get Started
Ready to improve your AI-assisted development? Start with Getting Started to learn the fundamentals, or jump to any section that interests you.
Built with evidence from dikuclient (63 PRs, 18 days), DikuMUD (165 PRs, 15 days), morpheum (76 PRs, 37 days), and diku (30 issues, 5 days).