AI gives different answers to the same question, and it does not remember previous sessions. This article explains how I use the term Harness Engineering for this site's AI operation practices.
How harness drift separates shared rules from tool-specific configuration, and how synchronization checks detect the mismatch earlier.
A three-part framework—repeatability, specificity, and verifiability—for deciding which recurring AI workflow problems belong in a harness.
How to write CLAUDE.md to give Claude Code project context, with before-and-after examples and the boundary between guidance and enforcement.
