Welcome to my corner of the internet. I document things here bugs, ideas, architecture wins, AI ramblings, time audits, and the occasional philosophical spiral.
If you really want something to work... keep it simple.
Everyone writes about ~/.claude. On Windows the real weight sat in %TEMP%\claude: one agent session, 117 GB, and 2,852 copies of the same 31 MB file. Where Claude Code writes, what the cleanup misses, and what's safe to delete.
A reverse prompt that flipped how I work with AI, why I use it as a memory-keeper and not a coach, and the small differences between models that actually matter.
Six months after publishing LLMR, the project has a proper home, a generator, and real numbers. Here's what changed and why it matters for anyone building with AI.
Twenty years in IT and I was still manually checking SEO dashboards like it was 2008. I don't have an SEO problem. I have a follow-through problem. So I automated the follow-through: GitHub Actions + Google Search Console API + Claude = 5 actionable issues in my repo every Friday at noon.
When exactly does this hook fire, how do you read submission data, and how do you avoid the most common traps? Three real copy-paste examples: Slack notifications, custom DB writes, and on-the-fly entity creation.
LLMR v2.0 is live; a JSON site index that gives AI systems your entire website in one fetch. Now with a formal spec, open-llmr.org, and a full ecosystem plan.
Two days, a dead SD card, an abandoned Acer BIOS, 83 million Mistral tokens, and a bot that only knew how to say "Task Completed". Building a personal AI assistant is not like the YouTube tutorials make it look.
From sneaking out of bed to watch MS-DOS boot, to building wind turbines in Haiti and architecting systems for Fortune 500 companies. This is who I am and how I got here.
I'm building Constituant, a platform where French citizens can vote on real legislation from the french National Assembly. Yes, anonymously. Yes, using AI summaries. And yes, I know exactly how broken that sounds.
A real debugging story: how AI assistants (GPT, Claude, Gemini) helped fix character collisions and movement in a Godot 3D project, turning a buggy scene into a functional game prototype.
Another legacy system integration nightmare featuring PowerPoint "documentation," Drupal hooks that fire too early, and web services that return completely different data than promised. When you can't test locally and debugging happens through client screenshots, every assumption becomes a liability. Spoiler: the eligibility check wasn't returning "Yes" or "No" like the specs claimed.
Why I switched from GPT to Claude for coding
After using GPT for a while, I found that Claude's approach to coding tasks was more aligned with my workflow.
Claude seemed to understand context better and provided more relevant suggestions.
Except now it doesn't anymore, and I'm not sure why.
Maybe it's the context window. Maybe I changed or I simply have a preference, but building a product that brings value is what matters.
And that's what they are doing, and that's what I am doing.