AIToday
Large Language ModelsAI Coding AssistantsZenn AI/MLPublished: Nov 9, 2026, 06:00 JST

Codex one-month trial: 139-file PR taught review-sized tasks

Send on LINE
Codex one-month trial: 139-file PR taught review-sized tasks

3 Key Points

  1. What happened

    Asking Codex to build a Rails blog produced a PR with 139 changed files and about 3,800 added lines, which the author says he could not fully follow.

  2. Why it matters

    He now splits work into reviewable units and keeps logs of what was requested and changed, to avoid what he calls understanding debt.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

The author came to Codex already comfortable writing Rails code himself and chose it because he was already a ChatGPT Plus subscriber. His first prompt was broad — build a blog, with a reference URL — and Codex generated the application, article list and detail pages, an admin screen, and tests, with Issue and PR text also written by the AI. The speed was real, but the gap between what the AI produced and what he understood became the central problem.

His response was procedural. He decided to check the AI's plan before implementation, split tasks into units he could review, and keep a log of what he asked for and what changed, a habit he traces to daily reports written during Fjord Boot Camp. He also read official documentation, listed unfamiliar terms such as context, skills, AGENTS.md, CLAUDE.md, agentic, and MCP, and read two books to organize the knowledge.

The Kamal incident shows why that discipline matters. A version-update thread led to a PR that passed RAILS_MASTER_KEY into Kamal's build settings, even though ReNovo runs on Render; leftover rails new configuration files had made Kamal look active. He removed the unnecessary Kamal settings while keeping what the Ruby update and Render build scripts needed, and concluded that a config file's presence says nothing about whether the technology is actually used. His stated goal now is to slow his judgment to a pace he can verify, rather than match the AI's output speed.

FAQ
Why did the author stop the first Codex-built blog app?
He had not conveyed his requirements well enough, so parts differed from what he intended. He paused development and moved to improving his own ReNovo app instead.
What went wrong with the Kamal configuration?
Codex changed settings assuming Kamal was in use, but ReNovo deploys on Render. Leftover Kamal files from rails new had made it look as if Kamal was in use.
What are AGENTS.md and Skills meant to do?
AGENTS.md tells Codex repository-specific rules and policies. Skills bundle the steps for tasks like creating Issues and PRs, so long prompts are not rewritten each time.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleAI firms have sold $300 billion of stock in 2026, with $100 billion more on tap