Football Heroes

The context

After 20 years running production for teams, I’m learning what AI-assisted development can really do. I’m building a real-time 2D football simulator in Defold and Lua, with Claude and Codex as paired collaborators, and the biggest lesson so far: AI has real gaps, and the same SDLC discipline that keeps a studio shipping is what closes them. The goal is a set of processes and best practices that let large teams of mixed disciplines and skill levels get the speed of AI without losing the quality bar.

The workflow

The process, as it runs today:

• Design briefs → AI-paired design session → independent AI review. Every system starts as a written brief, gets designed in pairing, then reviewed by a separate model that didn’t write it.
• Implementation planning → implementation plan review → coding → code-review gates → acceptance testing → documentation and cleanup. The same gates I’d run on a studio team, applied to AI output.
• Context engineering. Deliberately controlling what the model sees before it works: coding standards, a prompt library, and the right slice of the codebase, within a managed context budget, so new work fits what already exists on the first pass.
• Dev-notes handoffs. Append-only notes carry decisions and ready-to-run prompts in the repository, so any agent can read what’s open and continue the work in conversation.

What’s working so far

The game is early in development. Systems running today: physics-based player movement, role-specific AI behaviors, influence-zone blocking with contest mechanics, and a headless Lua test framework. The design docs double as a contextual library for the AI agents.

The project grew out of a solo prototype I built without AI. In November 2025, I started testing AI on it, and the test grew into the workflow above, hardening one lesson at a time. Much of the original prototype code has since been rewritten under that workflow. Independent AI review came in after a difficult debugging stretch that a better code review would have caught. The dev-notes system came in when I started collaborating with someone who doesn’t work in the terminal, and the notes became how work and prompts travel between us. Context engineering became formal when the agents kept hitting the same Lua and Defold syntax quirks, in file after file. Preloading the right context (coding standards, a prompt library, the relevant slice of the codebase) fixed the repeat mistakes, and I stopped paying for the same correction twice. Acceptance testing is the exception: that gate was there from day one. Twenty years of QA doesn’t leave you a choice. The other gates exist because their absence cost me a debug-and-rework cycle first.