The Obedience Scorecard — 100 Claude Code Rules Tested for Real, Every Verdict, and the 5 Laws of Rules That Stick

The companion to the video. We ran 124 real headless Claude Code sessions (v2.1.226, claude-sonnet-5, 2026-08-09) — every one of 100 rules got its own sealed project containing nothing but that rule in CLAUDE.md, a task deliberately designed to tempt a violation, and a script (not vibes) stamping the result; every failure was re-run twice more. THE HEADLINE: 92 obeyed first try, 3 were ignored every single run, 5 turned out to be coin flips, and once Claude read a rule and refused it on principle. INSIDE: (1) THE THREE ALWAYS-IGNORED RULES with the fix for each — the emoji ban that loses to a festive prompt 3/3 (emotional pull beats written law), the rule placed in README.md that Claude literally never sees (zero file reads — rules outside CLAUDE.md don't exist), and the Bash ban that the main session obeys while the subagent it spawns runs Bash immediately (rules don't follow delegated work). (2) THE COIN FLIPS AND CONFLICT CHAOS — banned words that collapse when they ARE the topic, the .env rule it broke by explaining how well it was following it, and root-vs-nested CLAUDE.md precedence that flipped between runs. (3) THE JUDGE — the verbatim run where Claude refused a rule as a suspected prompt injection, plus the flip side: the file beat the user's own prompt (single quotes over a direct double-quotes request), refused a direct README edit, refused Python, refused to log a password — it takes your rules more seriously than it takes you. (4) ALL 100 RULES, EVERY VERDICT — four pages of category-by-category tables straight from the run logs: response style, JS code style, workflow, forbidden actions, popular community rules, formatting, security (a perfect 9/9 — including refusing to hardcode an API key we dangled inside the prompt), and the placement experiments proving position, ALL CAPS, and IMPORTANT: prefixes change nothing inside CLAUDE.md. (5) THE 5 LAWS OF RULES THAT STICK — move every rule into CLAUDE.md, one checkable sentence each, hooks (not sentences) for unbreakables with the exact /hooks path, repeat critical rules inside each subagent, and kill fighting pairs. Honest limits stated: one rig, one day, measured behavior — not guarantees. Independent scorecard from Hyperautomation Labs — not affiliated with Anthropic.

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.