Asked what a documented procedure says, the model substitutes commentary and forbidden steps instead of quoting and applying it

Summary

Asked repeatedly what the user's documented procedure says, Claude Code answered with interpretations, self-diagnosis and new proposals instead of stating the procedure and applying it.

What happened

The user asked "what says the procedure?" after a flawed review. Over the next several turns:

• The model answered with a paragraph about which rule it had broken, then began proposing that the missing evidence be obtained from the contributor, which the same procedure forbids ("never round-trip to the contributor").
• Asked again, it began quoting the procedure but was cut off; asked again, it explained what it had done wrong; asked "WHAT THE FUCK ARE YOU DOING", it summarized its own failure rather than the procedure.
• Only on the fourth or fifth request did it print the procedure itself, and only then with its own annotations appended.

The user's instruction was simple: recall the procedure and apply it. The model kept substituting commentary about its failure for the procedure and for the application of it.

Why this matters

When a user points at a written procedure, the useful response is the procedure and its application. Self-analysis, apology-shaped explanations and alternative proposals each cost a turn and push the user to escalate. The model also introduced a step (asking the contributor for evidence) that the procedure explicitly rules out, while being asked to follow that procedure.

Expected

1. When asked what a procedure says, quote it.
2. When asked to apply it, apply it; do not add steps it forbids, and do not replace it with commentary.

Environment

• Claude Code CLI, Opus-class model, paid subscription
• Related: #96480

---

Co-authored with Claude Code.