Claude Code harness: start with CLAUDE.md and Skills only

Miyatchi's framing separates two ideas that are easy to confuse. Knowledge is what the AI is given — judgment criteria, rules, policies, and procedure documents kept as markdown files in the work folder, which he calls an "AI company" with departments for strategy, marketing, sales, and finance. The harness is how that content is made to take effect: when it is loaded, and whether the AI can decide to follow it.

That distinction drives the recommended order. CLAUDE.md is read in full at the start of every session, so it is the place for rules that should apply everywhere, and the official documentation suggests keeping it under 200 lines, moving job-specific content to Skills as it grows. Skills load only their name and description normally, with the body read when the related work comes up or when invoked by name, which makes them the natural home for a settled procedure.

Miyatchi also places the harness as the sixth of seven steps, after knowledge is organized and tools are connected, but before routines that run on a schedule. His own routine generates a YouTube report every Sunday at 8:00, and an access analysis for his company website was turned into a routine only after the method and report format were settled. The same logic, he argues, applies to subagents and hooks: adding them before the work pattern is clear risks stopping work that should not be stopped or cutting off context the main conversation still needs.