I compared five open embedding models to filter keyword lists down to a single niche. The top scorer wasn't the model I shipped, one wrong threshold made the best model look worthless, and the largest gains came from things the encoder never sees.
There's a recognizable look to AI-generated conference slides — the rounded teal cards, the icon-in-a-circle, the drop shadows. I took one of my own talks, had Claude reverse-engineer its design system into a reusable skill, and tested whether the skill could regenerate a deck in my style instead of the default one.
How I trained an 8B Llama on my own cardiology papers, from extracting prose out of Word docs, to a first attempt that failed because I trained on single sentences, to a working model that drafts abstracts, methods, and discussions in my style on a Mac.
How I traced Mixbook's animated Movie feature to its data API, found that the browser builds the video instead of storing it, and rebuilt a downloadable MP4 from the raw photos and music with crossfades, Ken Burns motion, and the original title text
Notes on adding multiple languages and right-to-left support to a Next.js app, covering logical CSS properties, locale formatting tags, LLM-assisted translation, and two hydration bugs to watch for
AI coding agents write a new migration file every time they change a table. Here's how a single pg_dump snapshot in the repo keeps Claude Code in sync with the current Supabase schema.
How to configure multiple Supabase MCP servers in Claude Code so you can work across different organizations and projects without constantly logging in and out