Workflow:

Built a meeting-intelligence pipeline that turns raw, error-prone transcripts into a structured, queryable knowledge base. Meetings get auto-transcribed by Krisp, whose speech recognition mangles the things I most need correct, like colleague names, customer names, internal product and architecture terms. I hand each transcript to Claude alongside a hand-built context document, and it works a fixed routine: read the context file, read the transcript, then reconcile every uncertain name or term against a master error table before drafting anything. Only the genuinely unresolvable handful surface as questions; everything else is corrected silently. Once I confirm those, it emits a cleanly formatted markdown summary in a manner I describe as a template: overview, topical notes, decisions, action items — and pushes the work items into Todoist so commitments don't get lost.

What makes it more than transcription cleanup is the back end and the feedback loop. Each summary hits Obsidian with YAML frontmatter and live Dataview queries, so open action items and meeting metadata behave like a database rather than static notes. In Cowork the whole accumulated Obsidian folder becomes fully queryable rather than merely searchable — instead of grep-ing for a keyword, I can ask questions that reason across months of meetings ("what were Todd's table-stakes asks, and has anything shipped against them"), with the model able to look across separate conversations. The other half is self-improvement: every clarification I resolve gets written back into the context document: its people directory, terminology glossary, and especially the ASR error table, so a garble I corrected once is corrected automatically from then on. Over time that one document has become a domain-tuned lens, and each meeting both draws on it and sharpens it, which is why the summaries keep getting tighter and need less of my intervention.

Beyond that: I use a Netatmo weather station which has a RESTful API (or sends to a cloud server that has one) - I pull that information (which I can see on the web and their apps) into my own VictoriaMetrics / Grafana set up on Kubernetes, via a Go app Claude built.

This app above was when I had my little aha moment: Netatmo's OAuth is slightly broken (issues with the different tokens and refresh). But I'd written a dog-ugly app which managed to work a while ago. Claude kept trying and strugglign to understand why its OAuth code wasn't working, and was asking me "are the credentials right?" etc. "Yup, I'm able to get data from my old app", then it said "If you have the source code to that app, I can figure out what's up", it looked, identified the issue, tried to work around it and then we "agreed" - "Hey, this should work this way, but it doesn't, and whether my old OAuth code should work or not, it does, so drop that in, and keep going". "Great, let's do that."