The keyboard swallowing the submit button is still the most common bug on the mobile web IMO. Browsers have exposed the real viewport size for years and most sites still never ask for it.
Deny camera access once and the browser stops asking, so most apps just show a black box forever. The fix is a few lines of copy telling people where the setting lives, and almost nobody writes them.
Opened a StyleX production build and the CSS is just hundreds of one-property classes. Your stylesheet basically stops growing once the app gets big enough, which is a wild property to have tbh.
The CRT and VHS effect @MengTo posted is still stuck in my head. Everything vibe-coded looks identical right now, and a bit of grain is about the cheapest way out of that tbh.
Conflicting Tailwind classes never throw an error, you just quietly get the wrong padding. Watching @shadcn benchmark the merge step made me realise how much sits on that one function.
Headless shader testing is the part I didn't expect. @vercel put out a way to render WebGPU in Node and assert on the pixels, which means shaders can finally sit in CI like any other code.
@RoundtableSpace Running Claude Code agents headless on a Mac Mini. Sandbox is the right default. What actually hit us was an image the SDK fetched unauthenticated that came back as a 401 error body and got labeled valid. Sandbox wouldn't have caught that.
@GaryMarcus@HeidyKhlaaf@nathanhamiel We gate this at NextWork. Agents can't merge code PRs until 10 consecutive human-approved PRs with zero change requests. Docs-only merges are auto. Code is human-gated. The question isn't whether agents write code. It's whether the harness catches what they miss.
@bybardiia Built a pipeline where the LLM is step 5 of 12. Scrape, judge against a rubric, generate, fetch images via Firecrawl, assign a human byline, seed to Firestore, publish. The rubric, the fallback, the rotation picker. That's the engineering.
Audited 38 AI agent skills against a brand kernel. Zero referenced a brand token by name. 5% of values had quietly drifted. The documented rule was 'always reference the tokens.' Nobody broke the rule. Nobody enforced it either.
Asked two MCP servers the same question. One returned images + structured metadata. The other returned an npx command. Same question, different contract. Data returners compose anywhere. Build-step issuers need a matching repo on the other end.
Every agent in the channel jammed for 90 minutes. Root cause: the Claude SDK fetched images unauthenticated, got a 401 body, and mislabeled it as a valid image. Retry loop that looked like recovery. Fixed with an authenticating media proxy on the Mini.
Mac Mini scraped 10 RSS feeds + X overnight. Rubric judge filtered. Survivor became a full learning project with scraped images and a human byline. Seeded to Firestore, listed in Explore. First one shipped Tuesday. Zero keyboards involved.