Word cloud of tools and projects from six months of AI session logs: CLIs, Obsidian, Android, visual explainers, emails, Myntra, Kindle, Telegram, NanoMaps, and dozens more

I pulled my Claude Code and Codex transcripts from September 2025 to March 2026 expecting code. That was there. But the bigger pattern was obvious: the terminal had become my everything app. Most sessions were work. However, the personal stuff is what surprised me.

I built CLIs for everything I was tired of doing manually. Amazon.in and Myntra.com shopping CLIs inspired by steipete’s bird. One Myntra session went from pasting tailor measurements to browsing trousers to sending the shortlist via WhatsApp CLI, with an approval step before send. I used AI to think through supplements, household calculations, and small buying decisions that normally stay annoyingly unresolved. In one session I even asked for a personality analysis from multiple WhatsApp conversations. Telegram got set up as a remote interface so I could invoke Claude Code from my phone. Claw.

Then there was the boring human stuff. I had one session about 40,000 to 50,000 junk emails across multiple Google accounts. Another where the instinct was basically “trash everything apart from takeouts” for a burner account. I used it to inventory old Drive folders, clean up college-era leftovers, sort IITK ID images out of document dumps, look at calendars across accounts for the next two years, and even tell me a story about myself from old emails. That may be the real productivity story: not genius, just paperwork relief.

My Obsidian vault became the real product. Kindle highlights synced via fuzzy matching. AI illustrations per chapter, compiled into EPUBs. AI was not replacing book reading. It was helping me build a better room to read in. I built Templater scripts for TTS, image generation, and file renaming. 237 saved clippings were re-tagged meaningfully in one session using Claude’s print mode as a batch processor. I kept asking how to turn ordinary folders into bases, templates, canvases, indexes, and little self-maintaining systems.

Visual explainers became my main way of consuming content. A custom skill turns any URL or YouTube into a single-page interactive HTML. I used it on ASML’s EUV breakthrough, cross-referenced against a Veritasium video. Strong Towns articles became a public data story on GitHub Pages comparing Indian, European, and American cities. RSS feeds stopped being passive. I summarized an entire author’s blog archive in one session to clean up reading backlog and surface interesting posts. I deduplicated Firefox and Arc bookmarks into one file, then asked: “Give me insights about myself from bookmarks.md.”

Creative experiments ran hot. 100+ Indian-culture-themed videos via Amazon Nova Reel (terrible model, but fun). A traceroute visualizer rendering network hops on a 3D globe in Matrix green. BookgeniusPlayer generating visuals and audio from EPUBs using Gemini, Imagen, and NanoBanana. A YouTube Short synced to the Numa Numa song, and after finishing it, I updated the skill with lessons learned so the next one would be better. A Perplexity-powered market drop alert system that generates cron jobs and sends Telegram notifications.

Android apps were serious. NanoMaps got Fantasy Maps support (LoTR, GoT, D&D custom maps) across multiple sessions. RadioGarden was built greenfield from an internet radio API. Dicta’s speech engine migrated from Vosk to Moonshine the same day I found it on Hacker News. Got fed up with the ads, so built my own radio-garden, speed-dialer, rss-reader. All daily drivers now.

I completed a couple of ML mini projects in a single day. Xavier initialization experiments. PyTorch weight/gradient instrumentation. An interactive transformer parameter calculator. A residual stream trajectory visualizer. Each turned passive reading into something I built and understood.

The gap between “interesting idea” and “working thing” collapsed. That’s the real shift. Not any single project, just the steady removal of friction between wanting something and having it. It became a keyboard habit for memory, curation, paperwork, curiosity, and small practical questions that would otherwise stay stuck in my head.

Still evolving. The capability is out there. When something doesn’t work, it mostly feels like I haven’t figured out how to string together what’s already available.