This month I pointed an AI coding assistant at one of my own blogs and asked it to tell me the truth about it. It pulled all 100 published posts through the WordPress REST API into local files, then counted things: 36 posts published on the same day, 19 of them between 110 and 127 words, one income figure repeated 108 times across the site, and 57 posts with no internal link at all. That job would have taken me a week by hand, and I would have stopped counting at about post twenty.
That is the honest shape of what these tools do for me: not writing clever code, but doing a large, boring, well-specified job without getting bored. Below is what I use daily, what I have not used and will not pretend to have used, and the review routine that keeps the generated work from quietly breaking something.
What this is based on: daily use of Claude Code on my own projects, on a paid consumer plan that includes it. GitHub Copilot and Cursor are not tested: I have never run either one, so everything I say about them comes from their own documentation and pricing pages, checked on 20 September 2026 and linked below.
The work I actually hand over
My projects are not a big engineering codebase. They are several small sites in different shapes, and that turns out to be the interesting case, because the assistant has to switch context constantly:
- Multi-page static HTML sites, where each page needs its own title, meta description, canonical and sitemap entry. Tedious, repetitive, and easy to get subtly wrong by hand.
- WordPress, both as a theme and as data: the bulk edits I described above go through the REST API, not the admin screen.
- A small Node service behind Caddy on a VPS, with systemd keeping it up. One process, one reverse proxy, one certificate, and a surprising number of ways for those three to disagree.
- One-off scripts. The measurements in my other posts come from a shell script I wrote with the assistant, which runs curl five times per site and prints a CSV.
The single biggest improvement to my results had nothing to do with which model I use. It was writing my rules down in a file the assistant reads at the start of every session. Most of them are about how I build sites, and I have printed them in the post about starting my studio. One is worth telling here because of how it was learned: never build a deployment archive with PowerShell’s Compress-Archive, because the paths it produces broke the CSS on a site I uploaded to a static host. Use Windows’ own tar.exe instead.
That last rule is worth dwelling on, because it is the exact shape of the failures you get with these tools. Nothing the assistant told me about zipping a folder was wrong. It simply did not know what the host would do with that particular archive, and it had no way of finding out. The code was fine and the result was broken, and only the person looking at the live site could tell.
Copilot and Cursor, strictly on paper
I have not used either. I am including them because they are the two tools people ask me about, and because a comparison that only contains the thing I already like is not a comparison. What follows is what their own pages say, nothing more.
| Tool | Individual plans as published | Have I used it? |
|---|---|---|
| Claude (includes Claude Code) | Free; Pro at $17/month with the annual subscription ($200 up front) or $20 billed monthly; Max from $100/month, sold as five or twenty times Pro’s usage per five-hour session | Yes, daily |
| GitHub Copilot | Free with 2,000 completions per month; Pro $10 per user/month with $15 of monthly credits; Pro+ $39; Max $100. Business and Enterprise have no price on that page | No. Not tested |
| Cursor | Hobby free with limited agent requests; Pro $20/month; Pro+ $60/month; Ultra $200/month. The page offers 20% off yearly billing without showing the annual figure | No. Not tested |
Sources, in order: Claude pricing, GitHub Copilot plans and Cursor pricing.
Two observations that do not require me to have used them. First, all three have moved from “a subscription” to “a subscription plus a credit allowance”, which means the monthly price is now a floor rather than a total: Copilot’s Pro plan quotes $15 of included credits on a $10 plan, and Cursor describes its tiers as multiples of “Pro limits” without publishing the underlying number. Second, they are not the same category of product. Copilot’s headline number is completions, which is autocomplete inside your editor. Cursor is an editor built around the model. Claude Code runs in a terminal and edits files across a project. If someone asks me which is best, my honest answer is that I only know what the third one is like to live with.
How I review what comes back
This is the part that decides whether these tools save time or cost it. My routine, in order:
- Read the diff, not the summary. The summary tells you what it meant to do. The diff tells you what it did. These are different documents more often than you would like.
- Check the files it touched that I did not mention. A change to one page that also edits the shared header is either exactly right or the start of a bad afternoon.
- Verify by measuring, not by asking. After a bulk edit, I do not ask whether it worked. I search the result for the strings that must not be there any more and expect zero matches, and I load the live URL myself.
- Check anything with a number in it. Prices, limits, dates, plan names. This is where confident wrong answers concentrate, and it is the reason every figure on this site carries a source and the date I checked it.
- Keep it away from anything I cannot undo. No publishing, no deleting, no touching a live database without a backup taken first.
Step three is the one people skip. Measuring my own sites on 20 September 2026, I found that none of the five sends an HSTS header and that my VPS does not advertise HTTP/3 even though the server software supports it, most likely because UDP port 443 is closed. Neither of those came up in any session with an assistant, because I never asked. The tool answers questions; it does not notice what is missing.
The certificate on that same VPS is the other reminder. On shared hosting the certificate renews itself and I have never had to think about it; on the machine I administer myself it has broken, and no assistant was going to notice that the renewal had failed before visitors did.
What I would tell someone starting out
If you are learning, the worst use of these tools is the one they are best at: producing a working file you do not understand. You get a site and no ability to fix it. Ask for explanations, write your own attempt first, and let the assistant review you rather than the other way round for the first few months.
If you already know roughly what you want and the job is large and repetitive, this is the best money I spend on software. Not because the code is brilliant, but because it does not lose patience halfway through a long, dull job. The parts that still need me are the parts that always needed me: knowing what the site is for, checking the numbers, and looking at the live page afterwards.
If you want the longer version of my experience with the one tool I do use every day, it is in my review of building my own sites with Claude. For everything that is not code, there is a short guide to which chatbot I would start with for which job. The setup these tools run on top of is described in my WordPress walkthrough for this blog, and how all of it fits into running a one-person operation is in my actual working setup.
Prices and plan details in this article were taken from the vendors’ own pricing pages on 20 September 2026 and are linked above. I am not in any affiliate programme, so the outbound links here carry no tracking and earn me nothing.