One Config, Parallel Agents, and Cursed Fonts

Manuel Matuzovic's slides go past the usual alt-text checklist: fixed headers can eat the viewport at high zoom, and `aria-haspopup="true"` tells assistive tech a navigation is a menu. The popover examples are useful too.

Two database queries can start together, then `yield*` an async generator to stream whichever result arrives first. That gets useful data to the client without making it wait for `Promise.all()` to finish.

The awkward case is one person splitting a block while another edits the text that moves. Notion's text-slice IDs and search labels help locate the edit after the split, a useful look at the production details behind CRDTs.

Animal Crossing's cosy island look—warm colors, pill shapes, chunky game-like buttons, and motion—gives this React component library a clear visual direction. I can see it making a small app feel delightfully odd.

Critical finds the CSS needed for above-the-fold content, inlines it, and defers the rest. A fast static path handles prerendered pages; Playwright is optional when a page needs a real browser render.

Bastardica combines fonts into one hybrid typeface, with plenty of room to make a readable face look subtly wrong. This is the sort of tool that can make a whole team question their eyes. Great idea for the April's fools day.

Superdiff compares nested objects and text, and can stream or read files for larger inputs. It gives you a structured diff to inspect instead of making you squint at two giant serialized JSON blobs.

I'm building this to keep one config for several AI agent harnesses and their providers. The idea is a single place to manage which tools and backends a project should use as the options keep multiplying.

Orca runs coding agents side by side in isolated worktrees, with terminals, diffs, a browser, and tasks in one app. The coordination layer looks more useful than another wrapper around a single model.

Prism ML lists a 5.95 GB language-model pack and reports 98.2% of FP16's score across 14 thinking benchmarks. The size is for the model weights alone; the score is the project's benchmark average, not a universal quality guarantee.