Note: This post is an experiment where I'm using AI to help craft the text. I'm working in VS Code with an AI agent (Claude Sonnet 4) that has access to my codebase and terminal output from the Zola server. As part of this exercise, everything is being changed through my voice – I haven't touched the text directly at all. I'm dictating via voice-to-text, so this isn't a transcript but my thoughts turned into readable text with AI assistance.
LEGO Me
I recently carved out some time to fix broken links throughout this blog's archive. To automate the work, I wrote a tool called wayback.py that scans for URLs and replaces any dead links with their Wayback Machine snapshots.

During that process I rediscovered the original avatar generator I used back in the day. It ran on Adobe Flash. That Flash-based tool was an unexpected bit of fun, and thanks to Ruffle it still works. I fired it up again and crafted a fresh version that you can see above, with the old original image below.

Here you can see my original blog post when I discovered this generator: my 2009 post on this blog. If you want to try it out yourself, here's a direct link to the file through the Wayback Machine: original Flash file.

And here's my current avatar that I've been using since 2009 – started with Twitter, then spread it everywhere. I've modified it quite a bit over the years: flattened it, removed the plastic glare, added a beard when I grew one. What you see now is more inspired by the original than actually being it. Funny thing is, I'd completely forgotten about the hands holding something until I saw the full image again after all these years.
Final Reflections
Dictating this to Claude has been an interesting experiment. It picked up my writing style reasonably well, but lost some personal touch – not quite using my usual phrases and words. It's also taken some fun out of writing, probably because I actually like to write.
For now, I'll stick to writing my posts myself, but I've already found LLMs useful as advanced spell checkers. I used Google Gemini Pro on my last post – rejected about 25% of suggestions because I preferred my style, but most caught actual English usage issues. Since I'm not the best English writer, that's genuinely helpful.
The bigger question is whether the method matters if the end result is the same. What's in my brain needs to transfer to text – keyboard or voice plus LLM, does it really matter? We aren't there yet, but LLMs improve constantly. When the quality matches what I'd write myself, maybe I'll try this again. Time will tell.