nomuraya — The Curious Operator Hands-on accounts of building, breaking, and rebuilding AI agents. urn:nomuraya:blogs:feed:en:v1 2026-06-17T08:10:24.662414+09:00 nomuraya (= shimajima / 中翔, the same person) Tests Pass, Design Breaks: Why TDD Can't Hold the Line on Design Intent urn:nomuraya:blogs:en:tdd-does-not-guarantee-design-intent 2026-06-17T08:10:24.284196+09:00 2026-06-17T08:10:24.284196+09:00 TDD verifies that the implementation matches the tests. It doesn't verify that the tests match the design intent. In the AI era, that gap silently widens. The Humans Become the Bottleneck: A Structural View of AI-Augmented Teams urn:nomuraya:blogs:en:human-bottleneck-ai-design 2026-06-17T08:10:24.283548+09:00 2026-06-17T08:10:24.283548+09:00 A strange thing happens in organizations that introduce AI agents: the AI gets faster, and the human side becomes the bottleneck. This is structural, not a skill issue. The fix isn't 'more humans' — it's redesigning where humans are required. I Copied a Google AI Studio Session by Hand. 68% of the Data Was Gone. urn:nomuraya:blogs:en:google-ai-studio-session-data-loss 2026-06-17T08:10:24.282875+09:00 2026-06-17T08:10:24.282875+09:00 I tried to preserve a long Google AI Studio session by copying the conversation manually. Measured against the JSON export, on average 6-7 out of every 10 KB was missing. Here is what disappeared and why. Building a 3D Scene from 30 Photos: Getting Gaussian Splatting Running on Colab urn:nomuraya:blogs:en:gaussian-splatting-2d-3d-getting-started 2026-06-17T08:10:24.282243+09:00 2026-06-17T08:10:24.282243+09:00 I wanted to touch Gaussian Splatting but kept stalling at the paper's math. So I went the other way: run it on Colab first, then read the paper. Here is the 30-minute path that worked, including the bits that broke. Why 'Just Be Careful Next Time' Never Reaches an AI urn:nomuraya:blogs:en:ai-hook-stop-points 2026-06-17T08:10:24.281427+09:00 2026-06-17T08:10:24.281427+09:00 AI doesn't remember 'I was told this yesterday' the way a human does. Telling it to 'be careful next time' assumes a memory it doesn't have. What works is making the constraint a physical hook, not a verbal instruction.