Claude Code 2ヶ月半: rules must sit outside the AI

The setup was three parts: a CLAUDE.md file holding purpose, prohibitions, price floors and writing standards; a ledger recording decisions, measured figures and failures, all dated; and a separate subagent acting as auditor that inspects client-bound writing and money decisions but does not implement anything.

One failure shows why the split mattered. When a blank-context subagent reviewed deliverables, it produced 5 findings, 2 of which had already been ruled on by the human, and one nearly became a fix that overrode a human decision without asking. What stopped it was not a more context-aware model but the ledger, which the main model had not consulted despite holding the context.

A second episode points the other way: after the human decided not to require measurement dates on external figures, the auditor objected, and two days later the human reversed their own decision and adopted the auditor's position. The operator also notes that line numbers cited from the ledger drifted four times in a single day — for example :1309 against an actual :1333 — because the ledger grows by inserting rows, shifting everything below.