https://b5874bfb.tim-o-dot-com.pages.dev/posts/ Recent content in Writing on Tim O'Brien Hugo en Fri, 17 Apr 2026 00:00:00 +0000 https://b5874bfb.tim-o-dot-com.pages.dev/posts/context-engineering/ Fri, 17 Apr 2026 00:00:00 +0000 https://b5874bfb.tim-o-dot-com.pages.dev/posts/context-engineering/ <div class="callout callout-tldr"> <span class="callout-label">TL;DR</span><p>LLMs are prediction engines that generate responses from the full context they&rsquo;re given. In older models (GPT-1 through 2025), you could improve performance by using prompting tricks to push them toward usable output (e.g.: personas, examples, &ldquo;don&rsquo;t do X&rdquo;). Frontier models are capable enough that those same tricks can push them too far, activating patterns you don&rsquo;t want.</p> <p>Lean, intentionally distilled context now beats elaborate prompts, especially when prompts include extraneous details that may invoke unwanted behavior.</p>