Today's posts highlight the gap between marketing promises and technical reality, especially when AI-generated code meets production standards. From Google's flawed "modern" guidance to practical CSS techniques and agent evaluation economics, these reads cut through hype with concrete examples.
Maybe Don't Rely on Google's "Modern Web Guidance"
Google's AI-powered Modern Web Guidance claims to generate accessible, performant code but falls short on all fronts when tested. The showcase accordion example ships with WCAG violations (missing focus indicators, broken keyboard navigation), Firefox-incompatible animations using discrete properties, and non-performant patterns despite explicit guidance existing for each issue. The core problem isn't just bugs—it's that LLMs are non-deterministic and can't reliably follow rules, making traditional pattern libraries far safer for production use.
Revealing Text With CSS letter-spacing
This technique animates letter-spacing from negative values (overlapping characters into illegibility) to positive spacing for reveal effects. The approach works particularly well for interactive states like checkbox labels and acronym expansions using existing pseudo-elements:
abbr::after {
content: " (" attr(title) ")";
letter-spacing: -1em;
transition: letter-spacing 0.3s;
}
abbr:hover::after {
letter-spacing: normal;
}While a hypothetical ::nth-letter selector would enable more granular control, these methods demonstrate what's achievable with current CSS.
The State of CSS Centering in 2026
Despite hundreds of centering methods floating around, only 10-15 are truly unique and valid in 2026. The key insight is understanding alignment fundamentals across layout contexts (Flexbox, Grid, Block) rather than memorizing tricks. Modern features like text-box: trim both cap alphabetic for optical centering and anchor-center for positioning are now viable:
.dialog {
position: absolute;
position-anchor: --trigger;
inset-area: center;
/* or: top: anchor-center; left: anchor-center; */
}The safe and unsafe alignment keywords prevent content overflow disasters, something that matters more as layouts become more dynamic.
Designing Efficient Verifiers for Legal Agents
LangChain and Harvey reduced legal agent verification costs by 60-1000x by switching from frontier models to DeepSeek V4 Flash and batching evaluation calls. Across 2,348 rubric criteria, DeepSeek matched Claude Opus performance while making reinforcement learning post-training economically viable. Targeted prompt tuning dropped false-pass rates from 15.6% to 14.2%, though even frontier models disagreed on 4-5% of labels, exposing inherent ambiguity in verification tasks.
Today's reads share a common thread: tools and techniques that look good in demos often break down under scrutiny. Whether it's AI-generated code failing basic accessibility tests, CSS centering cargo cults, or expensive model overkill in agent evaluation, the most valuable engineering skill in 2026 remains critical testing and understanding fundamentals over trusting abstractions.