Adrian Szabłowski

Writing

Notes on shipping AI features and the gap between what a model builds and what you meant.

2026

Jul 24 · 9 minLost in the middle is architectural. Reordering your prompt won't fix it.A March 2026 paper finds the U-shaped curve in untrained transformers, before a single training step. That changes which lever is worth pulling.

Jul 16 · 6 minClaude Code context bloat is MCP, not your promptA logging proxy reported 93,780 input tokens on one Claude Code request. The API said 2. Measured properly, MCP was 43k of my 99.1k context, not the prompt.

Jun 30 · 6 minNested Suspense: control the order your content revealsNested Suspense decides the order your sections appear, not when their data is fetched, and it will not save you from a waterfall.

Jun 25 · 10 minWhy AI builds the plausible feature, not yoursAI coding tools predict the most plausible version of a feature, not the correct one for your product, because they read only your context and your code.

Jun 22 · 15 minYour AI isn't too weak. Your evals are missing.A cheap eval harness for an LLM feature found a bug no bigger model could fix, exposed one of its own checks as wrong, and turned model choice into hard data.

Jun 16 · 6 minWhat React Doctor caught in my portfolio, and what I ignoredI ran a static audit over my own portfolio the way I would over production code. It flagged twenty things, and the real work was deciding which to overrule.