Studies and working notes by Jonathan Simone, with methods, results and supporting records. Each article links its data or sources.
Teaching an AI agent a project's rules.
Given only the task, agents passed 0 of 38 house-rule checks on a synthetic project; with source-linked library tools they passed 34 to 38 of 38. On 37 real-repository tasks (222 builds), a compiled guide (48.8%) and library tools (49.0%) did not beat the ordinary workflow (53.8%).
What holds up when we test published Voynich statistical claims, paleographic transcriptions, scan-aligned word geometry, and candidate production mechanisms across held-out quires.
I directed AI coding agents to rebuild Baldies, a 1995 god game, without its source code. The Windows desktop build reads all 100 original levels from an owned copy; a procedural edition runs on this site.
Define the task, limit access, review what can execute and prepare a recoverable checkpoint before leaving an AI coding agent working unattended.
An author-run study of 1,979 commits linked to pull requests in one repository, examining changed files, recorded author identities and review records.
Single-workstation troubleshooting records for packaged desktop applications on Windows 11.
On one Windows 11 PC, ChatGPT Desktop launch results changed when two Windows permission entries were added, removed and restored.
On one Windows 11 PC, Claude Desktop launch results and Event 208 changed when two permission entries were removed and restored.