<p><strong>TLDR:</strong> Two posts today make the same point from different angles — AI tools are consistently bad at frontend work, partly because they’re trained on stale data and can’t see what they’re rendering. A separate piece argues that the real fix isn’t better AI, it’s better specs before you prompt. Meanwhile, a nostalgic but practical case for idiomatic design is worth reading if you’ve ever shipped a UI that felt like it was invented from scratch.</p>
<hr>
<h2>Why AI underperforms on frontend</h2>
<p>One post lays out four reasons AI coding tools keep failing at frontend: training data is outdated and template-heavy, LLMs can’t render or visually perceive output, they don’t understand architectural intent, and the browser environment is too chaotic to reason about statically. A separate post on React Native makes this concrete — AI assistants still generate multi-line shadow props and manual flex margins when <code>box-shadow</code> and <code>gap</code> have been available natively for a while. The author ships an agent rule you can drop into Cursor to patch the worst offenders.</p>
<h2>Spec-first prompting as a partial fix</h2>
<p>The UIPotion post argues that messy AI-generated UI is mostly a prompting problem. Write a structured component spec first — states, rules, accessibility requirements, output format — then hand it to the model. The author claims this reduces guesswork and makes outputs consistent across iterations. The site catalogs these specs as JSON you can paste directly into Cursor, v0, or Lovable. It’s a reasonable workflow, though it doesn’t address the deeper rendering and perception gaps the other post identifies.</p>
<h2>The case for idiomatic design</h2>
<p>A post arguing for reviving idiomatic design is more substantive than the nostalgic framing suggests. The core diagnosis: the mobile transition forced awkward desktop/mobile hybrids, and frontend frameworks let teams bypass HTML idioms entirely, so every product now reinvents its own interaction patterns. The eight practical rules at the end — use semantic HTML, prefer words over icons, follow platform conventions — are worth bookmarking. Apple and Substack get cited as modern examples that actually enforce strong design systems.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Hugo llms.txt module:</strong> A step-by-step guide to adding <code>llms.txt</code> and <code>llms-full.txt</code> output formats to Hugo sites, with include/exclude rules to control what AI crawlers index.</li>
<li><strong>Shadcn templates roundup:</strong> 15+ Shadcn UI templates for Next.js evaluated on App Router compatibility and production-readiness; Tailwindadmin and MaterialM flagged for dashboards.</li>
<li><strong>kiki static site builder:</strong> A self-hosted homepage/wiki builder in pure PHP, no JS, no database, under 100KB — ships with RSS, gopher hole support, and screen-reader-friendly output; free with attribution or ~$11 for the full version.</li>
</ul>


Frontend development news, CSS techniques, framework updates, and browser API changes.

Web Digest

AI coding tools consistently underperform on frontend work due to stale training data, inability to visually render output, lack of architectural understanding, and the chaotic browser environment. A concrete React Native example shows AI still generating outdated shadow and flex patterns when modern CSS equivalents exist. A spec-first prompting workflow is proposed as a partial fix — writing structured component specs with states, rules, and accessibility requirements before prompting. Separately, a case for idiomatic design argues that mobile transitions and frontend frameworks caused teams to abandon HTML idioms, with eight practical rules for restoring platform-native interaction patterns. Additional items cover a Hugo llms.txt module, a Shadcn UI template roundup for Next.js, and a lightweight PHP static site builder.

AI keeps shipping outdated frontend code, idiomatic design makes a comeback