I make games with coding agents.
I am a vibe coder, not a software engineer. I do not usually know whether a diff is good by reading it line by line. I know from the outside: the game moves, the tool runs, the benchmark looks sane, or the thing feels less broken than it did before.
That is a strange place to build developer tools from. It is also the point. I trust AI unconditionally. I distrust its process and output just as hard. The thing can exist and still need to be smaller, clearer, and better.
| Repo | What It Is For |
|---|---|
coding-agent-guardrails |
Monorepo for my agent-side guardrails: keep edits small, record what happened, and make PR handoffs visible. |
Most of this is side-project work around games. The guardrails are side projects too: tools I built because my own workflow is too loose without them. Some repos are usable tools. Some are experiments. Some are proof that I pressed the gas first and went looking for brakes afterward.
The pattern I keep coming back to:
- Build the thing with the agent.
- Keep the change small enough to inspect from the outside.
- Record what happened.
- Make the handoff visible.
- Ask whether the result can be better.
The work here is uneven on purpose. I am testing what helps when a coding agent can produce more code than I can fully understand. The answer, so far, is not blind trust or pretending to be an engineer. It is smaller loops, better records, and one more pass.