A solo app builder says OpenAI’s Codex writes most of his code, but his real productivity gains come from the rules he has built around it. His account offers a practical lesson for anyone using coding agents: keep decisions recorded, make design checks enforceable and require evidence before accepting a change.
OpenAI Watch analysis
What happened
In a HackerNoon article published on 12 November, Sam Brovda describes building a consumer mobile app with Codex. He says the project has produced dozens of internal test builds and that its latest full check ran about 560 automated tests. He is careful not to treat commits or lines of code as proof of productivity: the app goes to internal testers, not production, and large changes can be difficult to review.
Brovda’s process is concrete. He records dated decisions in an AGENTS.md file, asks Codex to build several labelled design concepts, and turns approved mock-ups into specifications the finished screen must match. Custom lint rules catch violations of shared design components. Each meaningful change must come with a written account of what ran, what passed and what was not checked. Releases require explicit approval.
He also describes a practical trap: Codex stops reading instruction files once their combined size reaches 32 KiB by default. His root file was about 44 KB, leaving some older, important rules beyond that limit. His fix is to prune superseded instructions, put standing rules first and split area-specific guidance into nested files.
Why it matters
A coding agent can make changes faster than a person can sensibly review them. Brovda says one commit touched 241 files, adding about 21,000 lines and removing 6,700. His contrast is a three-file bug fix he could understand. Speed is useful; a mountain of code that nobody can check is a rather less charming side effect.
His answer is not to read every line or trust the agent’s confidence. It is to make the work easier to inspect: keep decisions current, enforce design rules in code, ask for evidence and stop releases unless a person authorises them. Those habits offer a useful starting point for teams experimenting with coding agents, even though one builder’s account cannot establish how well the same setup works across other projects.
Our read
The strongest point here is not the headline-sized commit count. It is the unglamorous operating discipline around the agent: Brovda makes it show what it did, what it tested and where it stopped. That is a better measure of a useful workflow than counting lines, which is still a popular way to make a spreadsheet look busy.
His account is an individual experience, not a controlled productivity study. But it gives readers specific practices to try, and a sharp reminder that an instruction file needs maintenance if the tool may stop reading it halfway through.
What to watch
- Whether Brovda’s smaller, more reviewable changes make it easier to catch defects.
- How he reorganises instructions to keep essential rules within Codex’s reading limit.
- Whether his evidence reports reveal gaps that ordinary code review would miss.
Discussion spark: For coding agents, which safeguard matters most in practice: smaller changes, enforceable design rules, written evidence reports or explicit release approval?
Sources and evidence
- Building an App With OpenAI Codex: What Works, What Breaks, and What I Learned – HackerNoon (12 November 2026, 06:19 UTC)
OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.