Claude Code user: CLAUDE.md tweaks are 'harness' work

The Zenn piece is written from the perspective of someone who had been improving his Claude Code setup without a name for it. He had added rules to CLAUDE.md, built skills, and used hooks to block dangerous edits, but could not explain what a 'harness' was when asked. His conclusion is that all of this is harness work, and he sorts it into four boxes: the loop, tools, context and control.

The fixes are then ranked by strength: asking the model in prose (CLAUDE.md, skills), watching with hooks that stop a bad action at the moment it happens, and removing the possibility of error through settings, lint and tests. Using his own examples, he shows that moving a warning from CLAUDE.md to a hook made it reliable even when the model forgets, though his first hook version also blocked ordinary commands and rejected the commit that fixed it. He also notes that OpenAI writes fixes into its lint error messages, calling this a good kind of prompt injection.

Beyond the ranking, he stresses timing (instructions are read at different moments), testing the same scenario before and after a fix with a subagent, and subtraction — Anthropic reportedly removed a handoff mechanism for Claude Sonnet 4.5 once Claude Opus 4.5 no longer showed the same habit. He cautions that most of the benchmark numbers come from parties with their own products, and that the Terminal-Bench paper itself concludes model choice usually matters more than the harness. Still, he argues the old habits of tests, lint and documentation now pay off more because the agent can read errors and retry until tests pass.