How to use Codex to build and repeatedly refactor a project
An author delivered a large holiday project with three coding agents, Markdown specs and repeated Gemini reviews, while reporting serious code-quality problems and a mixed overall experience.
news.ycombinator.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Implement [project scope] from [Markdown specifications]. Review the codebase for duplicated effort and convoluted patches, then propose prioritized refactors. Apply approved changes and report remaining maintainability problems alongside delivered features.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source poster / publisher
- dudeinhawaii
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- The author delivered a feature-rich project in five days but judged its code poor and development unsustainable, despite improvements from repeated review and refactoring.
- Human involvement
- The author wrote specifications, spotted duplicated and convoluted code, and initiated full-codebase reviews and repeated refactoring.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Write Markdown specification files for the holiday project.
- Develop the project across three agents and many sessions.
- Send the full codebase to Gemini-3 Pro for improvement suggestions.
The author reports building a roughly 100,000-line C#, Python and PowerShell project over five holiday days using three agents. They wrote many Markdown specification files and describe about 80 sessions, including extended refactors using codex-max. The delivered result impressed coworkers and contained five times as many features as the author says they would have implemented alone.
Code quality remained a substantial problem. The author observed duplicated effort and convoluted patches, and periodically sent the entire codebase to Gemini-3 Pro for improvement suggestions. Large reviews and repeated refactoring reduced some of those problems. Even so, they describe the code as terrible and say they would not publish it on GitHub.
They call this particular project a net positive because it had no commercial value or customer risk. The same post contrasts it with a commercial-project resource leak introduced during another agent-assisted change, which cost a week of troubleshooting. The author's overall assessment is therefore mixed rather than an endorsement of unreviewed output.
Who it's for: Developers evaluating feature delivery against the maintenance burden of agent-generated code.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Hacker News · dudeinhawaii
Frequently asked questions
How did the author address poor generated code?
They repeatedly sent the entire codebase to Gemini-3 Pro for improvement suggestions and used large reviews and ongoing refactoring to reduce quality problems.
Was the delivered project considered maintainable?
The author says the task was accomplished but not sustainably, describes the code as terrible and says they would not publish it on GitHub.
Did the author consider the experience successful?
They call the holiday project a net positive with no customer risk, but describe their overall experience as negative after a separate commercial bug caused a week of lost effort.
More details
Remaining steps
- Repeatedly refactor the code using large reviews before delivering the project.
About this account
Three agents contributed implementation across many sessions, while the author supplied specifications and repeatedly directed reviews and refactors.
Platform: Hacker News. Attribution: https://news.ycombinator.com/item?id=46770857. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: dudeinhawaii on Hacker News
Source published
Added to Agents In Use
Source checked
Tools and topics
codex-max · Gemini-3 Pro · Markdown · C# · Python · PowerShell
Code review · Refactoring · Multi-agent coding
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.