Nolan Lawson says Claude now answers frontend performance questions that once sustained web-dev education

Programming now often means configuring an AI and its runtime to get a result instead of directly manipulating code. I hate that shift.

I do not think prompts are like C source code. C has defined semantics and can be checked in and expected to work later, while a prompt alone will not reliably regenerate the same application next week.

I set up my wife with a DeepSeek key and Flash-V4 image detection for a cash-poor biotech startup. She can feed it the current site, request changes, tweak the design, and even alter product text inside an image for about $20 instead of hiring contractors.

Decades of web experience give me an advantage because I can spot slop, judge performance, and direct AI iterations. If cheaper sites free a biotech company to spend on biotech rather than web design, I welcome that change.

I helped my mother generate a single static brochure page with ChatGPT, and she sent me a ZIP a few days later that I could upload. The free-tier reset limits took most of the time.

Your wife did not do this from zero. You supplied the DeepSeek setup and the practical knowledge of keys, folders, index.html, and what to ask for.

Doing it yourself still consumes your time, and somebody has to take responsibility when it breaks. That may be fine for a small shop, but a mission-critical site needs dependable support.

I am less motivated to write internal documentation because Claude can answer a reader's specific questions from the codebase, while explanatory documents go stale. I still see value in documents that record intentions and decisions, such as why we chose mutable structures or plan a Rust migration.

Documentation should capture why a system exists and why choices were made, rather than merely explain how the current code works.

Frontend work is relatively safe to hand to an agent because the result is visible and can be rewritten. A database can silently corrupt data, hide security bugs, or create a query language that becomes hard to change.

AI-generated frontend code often puts data logic in the client that belongs behind secure backend APIs. In that case, the frontend becomes a serious risk.

Nobody knows how large AI will be in three to five years, and claims that it will inevitably get bigger are speculation. We may already have plateaued.

Internet search by an LLM can poison its context. A search for something like a rental API may make one service appear to be the only definitive option.

I do not understand why Cursor and Viget moved from Solid and Lit to React when better-performing frontend options exist. I wonder whether an agent really needs substantially more tokens to write Lit, plain HTML, or HTMX.