Journal

Notes on building with Code from Spec — progress reports, lessons learned, roadblocks, and future directions.

• Confinement, broken and repaired — August 7, 2026 — The fix for anchoring gave the agent a map of its neighbors in the spec tree, and a more capable model used that map to break confinement.
• Could we stop reviewing code? — July 26, 2026 — I wanted the repo to keep only the specs, with the code generated by the build pipeline and nobody reading a diff. The theory this project just published says why that does not work.
• Growing pains — July 19, 2026 — The theory essay gave the project a vocabulary for analyzing its own promises — and the first thing it analyzed was how incomplete they were.
• Telling the agent what changed — July 1, 2026 — The fix for anchoring: a cache, a hash comparison, and a one-word disposition per spec position. What worked, and the elegant hypothesis that didn't.
• Anchoring on old code — June 16, 2026 — When regenerating code from specs, should the agent see its previous output? On the tension between codebase stability and silent anchoring.
• Confinement — June 10, 2026 — Why Code from Spec's generation agents see only the spec chain — and why taking context away from an AI makes its output trustworthy.
• Context management by construction — June 8, 2026 — How Code from Spec gives AI agents exactly the context they need — no more, no less.