DEV Community

王泉伟
王泉伟

Posted on

I Let an AI Agent Fix My Website: 7 Minutes, 4 Tasks, Honest Results

I run an AI tools directory — 450+ tools, a prompt library, and tutorials. After a while, the problems piled up: broken search-engine push, thin tool pages, a dead data collector. I wrote them all down into a 22-item fix list, then handed the 4 most urgent items to Codex, OpenAI's coding agent.

No hype in this post. Just what happened, including the two times it failed before it even started.

The task

The four P0 items:

  1. Search-engine push was broken — new content wasn't being pushed.
  2. Both sites were missing llms.txt files.
  3. Tool detail pages were thin — only 27% of listings had complete data.
  4. The data collector had been stalled for 8 days straight.

My rules for Codex were simple: the source code is on the server, not in your hands — don't pretend you have it. Don't touch the live site. And don't make things up — verify what you can, and say clearly what you can't.

It failed twice before it started

First attempt: Codex wouldn't even enter the working directory. A trust check blocked it, and it sat there for 18 minutes. The problem was my launch setup, not the agent.

Second attempt: it started, then stopped — a missing component (code-mode-host) meant it couldn't read files at all. I had to download the matching component from the official release, install it next to the Codex binary, and restart.

Total time spent on environment setup: over half an hour. Actual task time: 7 minutes.

Lesson one: an AI coding tool is only as usable as its environment. Permissions, components, and working-directory setup decide whether it can work at all.

What it did in 7 minutes

Once running, Codex was disciplined: verify first, then act, then self-check.

Verify. It checked both sites' /llms.txt URLs itself and confirmed the 404s. It opened a tool detail page and confirmed the template boilerplate was still there. For the push and collector backends, it couldn't log in — and it said so explicitly instead of pretending it had verified them.

Deliver. It produced two ready-to-deploy llms.txt files (I byte-compared them against my drafts — identical), a full report, a cron wrapper, and an operations runbook. Two PHP snippets were marked "needs adaptation" because the real source code wasn't available to it. It didn't claim they were drop-in fixes.

Self-check. It ran checks on its own output: byte-level comparison of the llms.txt files, syntax checks, timeout and lock tests for the cron wrapper. The PHP snippets couldn't be run — no PHP runtime on the machine — and the report says that plainly.

What it did NOT do

Three honest limits:

  1. No source code, no direct fix. The PHP patches are generic implementations. A developer still has to adapt them to the real codebase.
  2. No backend access, no backend answers. Why did push fail? Why did the collector stall? Those need a human in the server panel.
  3. No runtime, no proof the code runs. Written code and working code are two different things.

So my score for this run: about 70% done, and it said so about the other 30% instead of faking it.

Who Codex is actually for

Best for people with a real project, clear requirements, and a working environment — developers fixing bugs, indie hackers maintaining a site, operators tidying up code. If you only have a problem description and no source code or test environment, start small: let it organize the problem, list what to prepare, and analyze causes first.

And one rule I won't drop: never treat it as an outsourced developer you don't need to review. AI does the heavy lifting; a human does the acceptance check.

Where to find the tools

I keep the directory at https://www.zcyai.cn — 450+ AI tools by use case, plus a prompt library and beginner tutorials. The account-management workspace I mentioned (one place for Codex, Claude, and other AI accounts) is at https://zh.zcyai.cn.

If you use Codex, what's the most concrete task you'd hand it? Tell me in the comments — the most upvoted one goes into my next experiment.

Top comments (0)