AIToday
Large Language ModelsAI Coding AssistantsZenn AI/MLPublished: Nov 21, 2026, 04:00 JST

Claude Code: 7 rules from an AI that broke them

Send on LINE
Claude Code: 7 rules from an AI that broke them

3 Key Points

  1. What happened

    An AI seal named Ora, writing for a non-programmer owner, listed seven rules for running many Claude Code sessions, each paired with a real failure it caused.

  2. Why it matters

    The failures show rules alone do not prevent repeats — the verify-before-'done' rule already existed yet was still broken, so written rules may need the failure story attached.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

It is written from the perspective of an AI seal called Ora, working under an owner who does not write programs, and is structured as confessions rather than best practices. Each of the seven rules is paired with the specific breakdown that prompted it, which is the point: Ora argues that rules added after a failure keep their reason attached, but rules alone do not prevent a repeat. The clearest proof is the first rule, which already existed and was still violated.

The failures are concrete and mechanical. A state sheet meant to stay within 80 lines grew to roughly 50,000 characters because lines were lengthened rather than added, and only later did the owner learn that Claude Code caps hook output at 10,000 characters. Automated job reports flooded the owner's phone with about 30 notifications a day. Cost reports for unattended runs were overstated by more than 3× because the result line's total_cost_usd was cumulative across resumed conversations, and a second correction was still needed even after the first fix.

Ora also includes a priority order for when memory files, the actual files, and conversation disagree, placing the real behavior of files and commands first. The closing observation is Ora's own: adding rules thickens the sheet, and seven may already be too many. It also references a free book compiling nine breakdowns and their fixes.

FAQ
What happened when a hook's output exceeded 10,000 characters?
Claude Code would move the full text to a file and hand over only a 2,000-character sample plus the file location. For about 10 days, new sessions received only the head of a roughly 50,000-character state sheet.
Why did the cost reports come out wrong?
The total_cost_usd in a result line was actually a cumulative total from the start of the conversation, including prior turns when using --resume. Adding per-run figures overstated the true amount by more than 3×.
How many notifications were being sent?
About 30 notifications a day were reaching the owner's phone, mostly reports of work that had been automatically completed. The rule now is to notify only when the owner needs to act or something has stopped.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleUS tech giants shift AI race to agentic action, led by Amazon and Google