I manage most of a WordPress site through ChatGPT on my phone. I describe what needs doing, ChatGPT creates GitHub issues, and Codex works on coding tasks from a VPS. WordPress still serves the site, but I rarely open its dashboard.
That wasn’t the plan when I started in April 2025. I’d worked with websites for years as an SEO professional, but I wasn’t a developer and had never built one from scratch. I wanted to see how far AI could take me.
I built a sports content site answering “where to watch” queries. Unlike a static website, it needed constantly updated fixtures, kickoff times and broadcasters. Getting it online was one experiment. Keeping it useful was another.
My personal website was designed for ChatGPT-driven editing from the start. This one began as a conventional WordPress site and gradually acquired the workflow I use today.
It started in WordPress
I registered a domain, bought hosting and installed WordPress with GeneratePress in April 2025. AI helped me write code, but I edited and published pages through the WordPress dashboard.
The challenge was changing schedule data. I built a child theme with PHP shortcodes that rendered fixtures from JSON files. Editorial content stayed in the page; changing schedules lived in structured data.
That was more maintainable than rewriting individual pages, but the JSON files still needed regular updates. An outdated schedule could make an otherwise correct page wrong.
So the question changed from can I build a site with AI? to how do I keep it accurate?
GitHub went from backup to source of truth
GitHub came in July 2025, months after the site launched. At first, I still edited code and pages through WordPress, cPanel or FTP. The repository was mainly a backup and a record of changes.
I tried PHP and WordPress scheduling for maintenance tasks, then gradually moved data collection and deployment into scripts and GitHub Actions. Fetching, formatting and validating schedule data didn’t need to happen inside the CMS.
The bigger change came in February 2026: I started storing page content as Markdown in Git rather than treating the WordPress editor as the source of truth.
The workflow became:
Edit in Git → validate → review PR → merge → publish to WordPress
Publishing through WP-CLI still sends the content to WordPress. But now Git retains the canonical files, tests can catch mistakes, and pull requests show exactly what changed.
More importantly, ChatGPT and Codex can work with the same versioned files without needing to operate the WordPress editor.
Automate predictable work, verify the results
Schedule updates mostly run without AI. Scripts fetch fixture data, normalize it, check it and deploy it. Some run daily; others run more often. Ordinary code is better suited to this repetitive work than asking an LLM to update fixtures.
But an automated deployment isn’t proof that the live site works. I added live checks, including mobile-browser QA for errors, broken layouts and unrendered shortcodes.
Those checks needed fixing too. One workflow flagged draft pages as broken production URLs. Automation can produce false alarms; it needs testing just like the code it monitors.
SEO reporting works on the same principle. Jobs collect Search Console and analytics data; I use ChatGPT to explore what changed and whether it matters. That approach also informed my analysis of how long new pages take to rank.
Scripts handle predictable data and checks. AI helps investigate and implement changes.
How I run the site from my phone
I usually start by describing a problem in ChatGPT: stale information, an improvement idea or an unexpected change in search traffic.
ChatGPT helps investigate, scopes the task and creates the GitHub issue for me. When coding is needed, I can ask it to dispatch the issue to Codex on my VPS. Codex changes files, runs tests and opens a pull request. I can review the result through ChatGPT before merging.
The loop looks like this:
Conversation → ChatGPT-created issue → Codex on VPS → tests → PR → review → publish → live check
GitHub holds the durable record: requirements, code changes, checks and decisions. The VPS means Codex and scheduled jobs can run without my laptop. Telegram notifies me when something finishes, fails or needs attention.
Routine fixture updates still run as scripts without involving an LLM. I mainly step in when there’s a problem to solve or a decision to make.
WordPress remains essential for rendering the site. What changed is how I operate it, not which CMS it uses.
What I learned
If I started again, I’d move content into Git sooner, keep less operational logic inside WordPress and build live verification alongside deployment.
I also wouldn’t begin with a large autonomous agent system. The most useful setup has been a collection of explicit workflows: scripts for repeatable jobs, issues for defined tasks, pull requests for review and checks for production.
I still decide whether an SEO opportunity is worth pursuing, whether an automation is useful and whether a change is safe to publish. Removing repetitive work makes more room for those decisions.
I started by asking whether AI could help a non-developer build a website. It could. The harder question was how to operate that website reliably afterward.
The answer wasn’t a better prompt or a more autonomous agent. It was giving AI a system of versioned files, defined tasks and checks to work within.
I started with manual WordPress updates. Now I can manage most of the site from my phone.