Blog

• Four Runs, Same Wrong Number: What Context Actually Buys a Coding Agent

  August 1, 2026

  296 scored trials of a headless coding agent on real tasks from my shipping products, across four context regimes and two models. Including the bugs I found in my own harness and the doc gaps the failures point at.

• My Terminal Congratulated Me for Clearing the Screen

  July 31, 2026

  Glissa told me a session finished when all I did was clear the scrollback. Three war stories from making Claude Code's own signals tell the truth about what six parallel sessions are actually doing.

• Teaching an AI to Parse Emails Once, Then Never Again

  May 19, 2026

  When an unknown email format shows up, call an LLM exactly once to generate a reusable extraction profile, then parse that format deterministically forever. Zero AI calls on the hot path.

• The 403 That Refused to Be a One-Line Fix

  March 10, 2026

  A failing deploy, a hand-edited Azure role shared by three services, and why the real fix was an IAM-as-code layer with a drift gate instead of a permissions patch.

• We Audited 24 READMEs by Hand. Then We Deleted the Process.

  January 22, 2026

  Twenty pull requests into a manual documentation cleanup, we abandoned the epic on purpose and replaced it with a linter and a required CI check.