>That is an interesting result. Which AI setup are you using and what >does it cost?
I'm using Claude Code, which is a terminal mode app. In this case I created a directory and initialized git in it, and then started "claude". I built this with the Opus 4.7 model running in "/effort max" mode. I gave it an initial prompt (in the repo) which was 60 lines long, a fairly long prompt, but I wanted to cover some fairly specific things. It then generated a "SPEC.md" (I was probably running initially in "plan" mode). I then had claude code go through all the phases and further specify them into "PHASE-*-SCOPING" documents, which I then had another model, gpt-5.5 high, review via the "codex" TUI app (OpenAI's equivalent of Claude Code). Then I'd have Claude turn that into a SPEC file for that phase. I'd then have it implement that phase. Then run codex to review the changes. Here's a natural point for me to "/compact" to compact the memory before starting a new phase. At times I'd make decisions about whether it seemed like another review round was useful. Once I got to phase 7, I decided to defer most of the items in there (hot config reload for example, seemed potentially risky and low reward), so I broke a few things off into a "phase 6.5". I did use a skill I have here called "codex-review" which does a code review of the last commit. I also used a "frontend-design" skill that's commonly available, I think Anthropic publishes that one, for the web UI. I did 2 or 3 different versions of the web UI, just to get it to feel a little less generic. All of this is in the repo, and pretty much every step is a commit, if you *REALLY* want to dig into it. As far as cost, my work pays $110/mo for the "Claude Pro 5x" plan, and most months I pay another $100-ish on top of that to get the "Pro 20x" upgrade ($200+tax). I pretty much do everything on the Opus 4.7 with "/effort max", because it works really great for me. But, it does chew through tokens, if you get the $20/mo account this setup will always be running out of tokens. They just recently increased the number of tokens you get at each level, so the $100 level is *VERY* usable. I run a lot of experiments to try different things just to get more experienced using the tooling, which is why I'm at the $200/mo level. For example: https://linsomniac.github.io/rally-xy/ Work also pays for the $20/mo ChatGPT account, which is what I use for the "codex reviews". >assistance. I have a program written to use GTK2, which will go away at If that's software that you do or would share the code for, I'd be happy to feed it to Claude to see what it can do, to give you an idea. I'd bet it could do a great job with it, one of my co-workers has gotten through a few years of backlog in the last couple months, cleaning up technical debt. Forward porting GTK code is likely fairly straightforward for AI because it can do builds and address the issues that pop up, which might be a large part of the port. On Wed, May 20, 2026 at 2:12 PM Charles Curley < [email protected]> wrote: > On Wed, 20 May 2026 13:15:02 -0600 > Sean Reifschneider <[email protected]> wrote: > > > Sounds like we both got started coding around the same time. ;-) > > I've been experimenting with the AI coding tools for around 3 years > > now, and I have found a use for the tooling: In around 40 hours of > > my attention I have something that manual coding would have > > taken ~9 months of focused effort. > > That is an interesting result. Which AI setup are you using and what > does it cost? > > I do have a project which might be a good candidate for some AI > assistance. I have a program written to use GTK2, which will go away at > some point. An AI trained in the various versions of GTK might be > useful for bringing that up to the latest version of GTK. > > > > > This contributor arrangement is an experiment, and we'll see how > > it goes. If it plays out that this project has suffered because of > > the lack contributions by senior coders such as yourself, that is a > > useful deliverable in my mind. > > Fair enough. Carry on. > > > > > I appreciate your feedback on the wording, I'll ponder that because > > my intention is to have this experiment be about AI generated > > software, so I may want to select different wording. > > Again, fair enough. Carry on. > > -- > Does anybody read signatures any more? > > https://charlescurley.com > https://charlescurley.com/blog/ > >
