I make games with coding agents.
I am a vibe coder. I rely on AI for the programming. I set the goals, question its choices, and ask for evidence that the result does what I asked.
I draw on basic ideas from industrial safety: designing hazards out, interlocks, and pointing and calling. They shape the questions I ask when working with coding agents.
The thing can exist and still need to be smaller, clearer, and better.
| Repo | What It Is For |
|---|---|
| Masters’ Nudge | Short structural feedback while a coding agent works, with published comparisons of results and costs. |
| arc-9 | One manually invoked engineering review skill for Codex, tracing causes and proposing concrete changes from task requirements and evidence. |
| Coding Agent Guardrails | Scope limits, trusted checks, and records of the final state for coding-agent work. |
Games are where this started. The tools grew out of problems I kept running into while working with coding agents.
Sometimes I end up building a tool for the agent that was supposed to be building the game.
I keep coming back to what happens after the first working version: whether the structure makes sense, whether the extra complexity is necessary, and whether an improvement is worth what it costs.
The questions I keep coming back to:
- What is the goal, and what is within scope?
- What is the smallest change that meets that goal?
- Which mistakes can we prevent through the design?
- What conditions must be met before an action is allowed?
- What needs to be checked and called out before moving on?
- Does it actually run and behave as intended?
- What improved, what got more complicated, and what did it cost?
- What should we record so we can check the result and pick up the work later?