I Used to Say ChatGPT Was Useless. Here Is How I Work Now.
I built sites on platforms for years and said ChatGPT was not good enough for real work. This year my own site moved into code, and the AI I use now answers to a chain of command. Here is what changed, and what it means for a client.
A year ago I would not have believed this sentence: on the day this post went up, the site you are reading took 33 changes between 9 AM and 6 PM, and one of the biggest batches came from an OpenAI model reviewing work that was built mostly with Claude.
For most of my career I built sites on WordPress, Wix, Squarespace and, most recently, Webflow. And I used to say ChatGPT was not good enough for real work, and that I would never use it. Here is what changed, what did not, and what it means if you hire me.
From a website builder to a code repository
This month I moved pfendermarketing.com off Webflow. The site now lives in a code repository and publishes through Vercel. Every page and every blog post is a plain file I own, and the repository is connected to the AI tools I work with.
Before the domain switched over, I checked all 77 addresses I could find for the old site. Every one now lands somewhere sensible, except pages that were already dead on Webflow.
What the move buys me is control. On a platform, you work inside what the platform allows: its templates, its editor, its way of handling a form or a redirect. In a repository, if I can describe a change, it can be made, and every change is recorded as its own entry that I can read line by line and undo.
Show the numbers
| Hour (Eastern) | Changes |
|---|---|
| 9 AM | 7 |
| 10 AM | 1 |
| 11 AM | 1 |
| 12 PM | 2 |
| 1 PM | 1 |
| 2 PM | 5 |
| 3 PM | 2 |
| 4 PM | 4 |
| 5 PM | 10 |
It is also not the same as becoming a software developer. I tried that with my own internal tools and decided it is not the job I want. Directing changes to my own site is.
And it is not the right answer for everyone. A website builder gives an owner a login and an editor they can use without me, and for a lot of small businesses that matters more than control. If Wix or Squarespace is working for you and you update it yourself, I would not move you.
Why I changed my mind about ChatGPT
My objection was never that the tools were useless. It was that I could not trust them on work that had to be right, and in client work almost everything has to be right.
That changed this summer. GPT-5.6 Sol was the first OpenAI model I trusted with that kind of work, and GPT-6 Astra, released in early September, has kept that going. Most of my work still runs on Anthropic's Claude, and I am not going to rank them. They leapfrog each other every few months.
I no longer rule a tool out because of what it could not do a year ago.
Every assistant, the same tools
This is the part that changed how I work the most. Claude, ChatGPT, Grok Bot and Viktor are all connected to the same tools: the site's code, my email and calendar, my documents, and the systems I run PMC on. They share the same context too, so I can start something in one, pick it up in another, and switch back without explaining it all again.
- Claude
- ChatGPT
- Grok Bot
- Viktor
- Site code
- Email and calendar
- Documents
- Business systems
So no tool is a silo. If one is better at a task this month, I use it for that task. If one gets something wrong, another can check it with the same information in front of it. In practice most of my work still runs on Claude, but nothing is locked to it.
One model is not a team
The bigger change is not any single model. It is how the work is organized.
Grok Bot is the AI assistant layer I run PMC on. Instead of one chat window that tries to do everything, it is a set of named agents that each own a lane: fulfillment, sales, operations, marketing, partners and code.
- Fulfillment
- Sales
- Operations
- Marketing
- Partners
- Code
The chain of command is simple. I talk to one agent, which acts as chief of staff. It routes work to the right lane and keeps the board of what is open, moving and stuck. And anything that leaves the company, whether that is an email, money being spent or something being published, waits at the gate until I release it.
That gate is the rule that matters most. Nothing goes out in my name, or a client's, because a model decided it was ready.
A second set of eyes
Because every assistant works from the same tools and context, checking one model's work with another is easy. That is how the review of this site happened.
This week I had GPT-6 Astra review this site, which was built mostly with Claude. It screened every post on the blog, read 13 of them in full, and came back with a list.
- Posts screened
- 25
- Read in full
- 13
- Factual corrections
- 8
- Posts taken down
- 1
Most of the eight corrections were platform facts that had changed or been overstated: what the delivery apps charge restaurants, what OpenAI has said about its search providers, how Google handles a visitor who declines cookies. One was a statistic quoted without the limits of the study behind it. Separately, the review flagged claims about my own work that did not match my records, and that is why one post came down.
One model checking another's work is not a guarantee. It catches a different set of mistakes than the model that made them, and this time it caught mine. The rule that came out of it is simple: a platform fact needs a primary source from the vendor itself, and a claim about my own work needs a record I can point to. If neither exists, the sentence goes.
What stays with me
The tools do more of the typing, the research and the first drafts than they did six months ago. What they do not do is decide.
What goes out, what gets spent, and what gets said in my name or yours is still my call. So is the judgment about which problem in a business is worth solving first. That part of the work has not moved, and I do not expect it to.
What this means if you hire me
Three things, in plain terms.
- Faster turnaround on the routine work. The time that saves goes into the parts that need judgment, not into producing more of the same.
- Nothing goes out in your name without your approval. That is the same gate I run on my own business.
- Tools picked for the job, not out of loyalty. If the right answer for your business is the platform you already have, that is what I will tell you.
If you want to see where your own site stands first, the free site scan takes up to two minutes and needs nothing but your web address.
Common questions
Why did you move your website off Webflow?
Control. In a code repository every page is a file I own, AI tools can work on it directly under my direction, and every change is recorded and can be reversed.
Should a small business move its website into a code repository?
Usually not. If you update your own site and it is doing its job, a website builder like Wix or Squarespace is often the better fit, because you can change it without calling anyone. A repository makes sense when someone is actively managing the site and the platform is what is holding it back.
Do you use ChatGPT or Claude?
All of them. Claude, ChatGPT, Grok Bot and Viktor are connected to the same tools and share the same context, so I can use any of them for any job and switch between them mid-task. Most of my work still runs on Claude, but nothing is locked to it. GPT-6 Astra is the one that reviewed this site and found corrections I needed to make.
What is an AI chain of command?
A setup where AI agents each own a defined lane of work and report through one coordinator, with a person at the top who approves anything that leaves the business. The point is that no agent can send, spend or publish on its own.
Will an AI tool ever send something in my name?
Not without your approval. Anything that goes out on a client's behalf waits for sign-off, the same way it does for my own business.