Two Years of AI Coding, Visualised
What 500 conversations with Claude and Gemini taught me about how I actually build software.
I recently exported every chat I'd ever had with Claude and Gemini and asked an AI agent to do something I'd been quietly curious about: read through all of them and tell me what I had actually been working on. The result is a single-page thematic summary - three A4 sheets that condense roughly 500 technology and coding conversations into a handful of recurring themes.
Looking at it laid out like that was unexpectedly clarifying.
The shape of the work
The biggest surprise wasn't what I'd worked on - it was how consistent the shape of the work has been. Across 17 months, five themes account for the bulk of every chat: HTML/CSS/Bootstrap, JavaScript, WordPress, Classic ASP / VBScript, and web design & animations. The pattern is unmistakable: lightweight, single-file web pages that can be pasted straight into a CMS or a learning environment, with just enough JavaScript to feel alive.
I clearly have a type.
From "AI as reference" to "AI as builder"
The volume tells its own story. In late 2024 and the first half of 2025, I used AI mostly as a smarter manual - a few questions a week about WordPress, CSS, or Classic ASP. Then, somewhere around August 2025, the volume jumped roughly tenfold and the prompts changed character. They stopped being "how does X work?" and started being "build me X."
That shift is the through-line of the entire export.
One project that ate a third of my chats
The export also revealed something I hadn't fully appreciated: a single project quietly consumed about one in every three tech conversations. ASP4 / ASP.py - a Python runtime that parses and executes legacy Classic ASP/VBScript code - turned out to be the spine of the last twelve months. Parsing, runtime objects, ADO emulation, AWS install scripts, even trademark research on the name. End-to-end, AI-assisted, written almost entirely in single iterative sessions.
Without the export, I would have estimated this at maybe 10% of my AI usage. The real number is closer to 30%.
The other lessons
A few quieter patterns also stood out:
- I almost never want a build step. Vanilla web tech, Bootstrap from a CDN, one HTML file. The summary calls this a "signature deliverable," and that's fair.
- I switch languages mid-prompt. Dutch and English blend constantly, occasionally French. The AI handles it without blinking; I keep forgetting that's not a given.
- Education quietly drives a lot of the design. Many of my "production" builds - interactive timelines, quizzes, memory games - started life as teaching artefacts for bachelor students.
- I prefer small, cheap infrastructure. PythonAnywhere, t-class EC2 instances, IIS on a Windows PC. The conversations keep returning to "what's the smallest thing that will work?"
Why I'm sharing this
Two reasons. First, if you're using AI assistants seriously, I'd recommend doing this exercise yourself. Both Claude and Google offer full data exports. Aggregating them is humbling - and far more useful than any productivity dashboard.
Second, the summary itself is a small artefact in the same style as everything else on it: a single HTML file, inline CSS, no dependencies, ready to print as three A4 pages. It generated itself, in a sense - which feels like a fitting conclusion to two years of this kind of work.
Have a look: → Two Years of AI Coding · Thematic Summary
