writing
Writing
Notes on code intelligence, developer tools, verification, and AI-assisted software engineering. Benchmarks over debates.
featuredWe benchmarked AI agents on dead code detection across 60+ runs and 14 real-world repositories using Claude Opus 4.6. Graph-enhanced agents achieved 94.1% F1 with 100% precision — 156x cheaper, 11x faster, 2x better than the baseline agent alone.
read the benchmark →-
July 26, 2026A monorepo is not a folder: Preserving the full history of Pain to Profit
I consolidated three connected projects without throwing away their commits, dropped work, tests, architecture, or evidence.
-
July 26, 2026Moving a window is not resizing it: what Beckon 0.1.6 fixed
Beckon 0.1.6 fixes unreliable macOS window layouts with verified Accessibility frame writes, deterministic tests, and real hardware round trips.
-
June 9, 2026First look at predictive intents for Hermes: Architecture should be benchmarked not debated
SQLite beat Kuzu and an LLM baseline in a benchmark for a predictive intent system. The article is about running the test, not the result.
-
May 12, 2026If you keep retyping a prompt, make a skill
A practical guide to Claude Code skills: what they are, how to write them, where they live, when not to use them, and how to package one as a plugin.
-
May 4, 2026We threw out the file watcher and used git instead
Why our CLI watches your repo with `git diff` and `.git/index`'s mtime instead of inotify, how we keep a 50,000-node code graph correct under live edits, and...
-
April 27, 2026Seven bridges, one question, and the start of graph theory
A walk through 18th-century Königsberg, an impossible puzzle, and the moment a Swiss mathematician turned a city into a graph.
-
April 24, 2026Your AI keeps giving you a different architecture. Here's how we fixed it.
Ask an AI to draw your architecture on Monday and Tuesday and you get two different maps. We found a cheap fix that gets the same map every time.
-
April 24, 2026I told Claude to double-check itself. It got 55% dumber.
Adding a verify step made my AI agent 55% dumber. Here's why asking a weaker tool to check a stronger one always backfires.
-
April 22, 2026Your AI agent is spending 90% of its tool calls on map-building
Why AI agents burn most of their tokens rediscovering a codebase, and how a small file next to each source file fixes it.
-
April 21, 2026Dear Vibecoder,
Your AI doesn't see your code as a system. It sees words. Here's what that costs you and what to do about it.
-
April 20, 2026Supermodel Public API Explainer
A walk through every public endpoint on the Supermodel API: what each graph represents, why we ship it as a primitive, and what you can build on top.
-
April 17, 2026Why we built Supermodel
Why agent-written code demands a graph model, and how Supermodel builds one from your repo.