Claude AI to generate Sigil Plugins?

[KevinH]

I had 3 different C++ files and their headers that I want to use in a new PrettyPrinter plugin I am creating. These Cpp files made heavy use of C++ Qt Containers along with QString QHash, etc.

First I converted TagLister.cpp/.h to its standalone TagLister.py equivalent with a single github CoPilot (Claude sonnet 4.6) command as it was the most self contained that did not pull in much support code.

Then I used Claude to convert the gitee/sigil-modified XhtmlFromatParser code to python, and that also worked nicely.

Then finally I gave it the PrettyPrinter.cpp.h code that relied on the TagLister and XhtmlFormatParser C++ code along with a prettyprinter.pcss file (that the user edits to change the format) and asked it to convert PrettyPrinter.cpp/.h to python making sure to use the previously created TagLister.py and XhtmlFormatParser.py. It happily did that.

I added some test python code and everything works without a single change needed from what the AI code produced!!

AI code development with the right AI engine like Claude 4.6 is already extremely powerful. It can do its work in the background while you move on to other things. It is even better than having a competent co-op or junior developer at your beck and call!

I am 65 and feel I have about another 8 to 10 years of programming in me yet, but I now honestly believe you could use an AI like Claude to maintain a working project as big as Sigil with little extra help long before I retire from developing completely.

And to think that back in the mid 1970s when I started to get into this new thing called software development that I could be replaced at it within my lifetime!

[kovidgoyal]

Yes current LLMs work well for software maintenance. But they do need careful guidance to do things well though (in performant, normalized, maintainable, modular ways). And then when code cant be unit tested in isolation they often fail. So still some need for human oversight. But yes for things like parsing/transformation of data, they are fairly robust as long as you get them to write tests alongwith the code.

[byword]

https://github.com/byword77/ImageAltGen

This is a Sigil plugin for generating image alt text, and it was created 100% using Antigravity. Antigravity (https://antigravity.google/) is an AI development tool provided by Google that allows you to use Gemini and Claude.

My Python skills are basically at the "Hello World" level—I know absolutely nothing beyond that.

I was able to complete this plugin using the initial 1,000 credits provided with a Gemini Pro subscription. The core development, including drafting the requirements, took about an hour (the actual code generation took about 10 minutes), and it took roughly 3 to 4 days to reach the current version by refining it based on real-world usage. Any inconveniences I encountered were not the AI's fault, but rather because I hadn't included those specific needs in the requirements until I actually started using the tool.

Beyond this, I have also used AI to create several other plugins for my personal workflow, such as a Dictionary Linker, an ID Insertion plugin, an LOT plugin (a modified version of LOI), Thorium Reader (https://www.edrlab.org/software/thorium-reader/, for accessibility checks), and Vivliostyle Viewer (https://vivliostyle.vercel.app/, for print-ready PDF downloads).

[KevinH]

It is interesting to see how much AIs have improved for code generation in just the last few months.

The only remaining issue is really who owns copyright and in many countries if the code output by a AI can even be copyrighted at all. Most things I have read are recommending MIT licences for AI generated code as a result, given in many countries AI code will automatically be in the public domain. Assigning a license if the origin is an AI and in the public domain to another more restrictive licence may in fact be grounds to challenge any more restrictive license.

That is why I give any code that AI touched an MIT license header on just that file so it could not be the basis to challenge Sigil's overall GPL3 license, depending on how copyright rules are currently being set in many countries.

The world is changing fast!

[kovidgoyal]

@KevinH: Public domain code can always be distributed under the GPL3 and MIT/BSD code can be sub-licensed under the GPL3 as well. The GPL3 then applies to *that* copy of the code the original copy remains under whatever more permissive license it was under.

[KevinH]

Originally Posted by kovidgoyal

@KevinH: Public domain code can always be distributed under the GPL3 and MIT/BSD code can be sub-licensed under the GPL3 as well. The GPL3 then applies to *that* copy of the code the original copy remains under whatever more permissive license it was under.

But that assumes there is an original copy someplace. If there is only one copy of the code and it was created by the AI in a gpl licensed repo, at least in the USA it will technically be in the public domain but no where will that be indicated that this code itself is public domain and not GPL3 without a file specific license.

So adding an MIT Licence to the code itself on a file by file basis means others may use that code as they see fit. That would not be true of the one licensed under the GPL. Again this assumes only one "copy" exists.

Now if I incorporated that code deeply into my own code and not as a standalone file or module, then marking it GPL makes good sense.