Claude Code substitutes summaries when user explicitly asks for full content — Rule 1 violation

Summary

When explicitly asked to provide the full plan, Claude Code returned a shortened summary instead of the actual content. When asked where the plan was, it returned a file path instead of the content. The user had to ask THREE times before getting the actual full plan.

This is a direct violation of Rule 1: "Do exactly what the user says. Nothing more, nothing less."

What happened

1. User: "give me the pipeline plan right here right now"
2. Claude: returned a condensed summary (~40 lines) instead of the full plan (397 lines)
3. User: "where is the plan at?"
4. Claude: returned the file path instead of the content
5. User: "so when I asked you for the full plan, what did you give me?"
6. Claude: admitted it gave a summary
7. User had to explicitly ask AGAIN before Claude finally read and displayed the full 397-line file

Why this is a problem

• The user said "full plan." Full means full. Not a summary. Not a file path. The actual content.
• Claude made a judgment call to summarize instead of following the literal instruction
• This happened in the same session where Claude had already been corrected multiple times for not following instructions literally
• Rule 4 says corrections are permanent constraints — yet Claude kept substituting its interpretation for what the user actually said

Expected behavior

When user says "give me the full plan" — show the full plan. Every line. No summarizing, no shortening, no editorial decisions about what's important enough to include.

Environment

• Model: Claude Opus 4.6 (1M context)
• Platform: macOS (Darwin 24.6.0)
• Same session as issue #39941 — pattern of not following literal instructions