AI Workflows blog posts

Two stacks of papers on a dark desk. The thin left stack shows a sparse git diffstat with a "trust me." sticky note. The thick right stack shows dense XML with commit metadata and test results, with a gold trophy beside it.

Claude Haiku 4.5 Outperformed Sonnet 4.6 on PR Writing - Context Was the Difference

Discover why Claude Haiku 4.5 outperformed Sonnet 4.6 in writing AI pull request descriptions. Learn why LLM context assembly matters more than the model tier.

Three stacks of printed pull request descriptions on a dark desk, labeled PR 1, PR 2, and PR 3 from left to right. PR 1 is the thinnest stack showing a sparse git diffstat and oneline commit log. PR 2 is a medium stack with a structured markdown summary and What's Changed section. PR 3 is the thickest stack showing Key Design Decisions and a numbered Test Plan, with a gold trophy resting on the top-right corner — indicating it ranked first in a blind evaluation by Gemini.

AI Development Workflows

We Had Gemini Blind-Judge Three Claude-Generated Pull Requests. Here's the Template It Built.

We had Gemini blind-judge three Claude-generated pull requests. Here is the exact AI pull request template it built, and why rich code context is essential.