AI makes building cheap. What's left is judgment: what to cut, what comes first, how it should feel. Skills keep that judgment in every project. Full post: https://t.co/10XIHieVet
Skip the giant project.md. It gets reread on every message and turns into noise. A shadcn boilerplate with theme tokens and a base layout gives Claude context through real code. Keep CLAUDE.md to a few lines.
AI made it fast to ship the same mistake on every screen. The upside: fix a handful of components and most of the audit clears. Full post: https://t.co/wMGgY4uXKC
Telling the AI "make it accessible" doesn't work. In one 2025 study it made output slightly worse. Feeding real checker results back to the model worked best. Give it rules, then give it errors.
A clickable prototype sold the redesign faster than any deck could. The result: about 40% faster reviewer onboarding and about 50% fewer clicks to approve an application.
Most client workshops feel productive and change nothing. Two days of sticky notes and journey maps, and the product looks exactly the same the week after.
Then test with 2 to 3 real users within a week. One 20 minute call, one task. On that project, a long step-by-step onboarding became one checkbox on the signup screen. We only found it because the first version existed.