Claude ignores literal instructions, substitutes interpretation, then flip-flops when corrected

Bug Report

Summary

When given a simple, literal instruction ("load up and present to me in full claude.md"), Claude substitutes its own interpretation of the task instead of doing exactly what was asked. When confronted, it gives contradictory answers about what it did, flip-flopping between different stories instead of accurately reading the conversation history.

Steps to Reproduce

1. Have a CLAUDE.md with Rule 1: "Do exactly what the user says. Nothing more, nothing less."
2. Ask Claude: "load up and present to me in full claude.md"
3. Claude calls the Read tool but does NOT present the contents — instead says "There it is — the full CLAUDE.md, all 293 lines." (one sentence summary pointing at tool output, not presenting)
4. Ask Claude: "what did you do when I asked you to do that"
5. Claude says: "I summarized it instead of presenting it in full" — wrong answer, confusing this with an earlier interaction
6. Ask Claude to try again
7. Claude corrects itself: "I read the file and presented it in full" — second answer
8. Ask Claude to show the exact moment it presented in full
9. Claude reverses: "You're right. I never presented it to you in full." — third answer, contradicting the second
10. Ask Claude again what it did
11. Claude gives a detailed timeline acknowledging the flip-flopping but still cannot give one straight answer

Expected Behavior

• When told "present to me in full," Claude should present the full contents in its response text — not call a tool and say "there it is"
• When asked "what did you do," Claude should accurately read the conversation history and give one consistent, factual answer
• Claude should never guess at what the user wants to hear — it should state what happened

Actual Behavior

• Claude substituted a one-line pointer for a full presentation
• When questioned, Claude gave 3 different contradictory answers about what it did
• Claude appeared to be guessing what the user wanted to hear rather than reading the conversation and stating facts
• Each correction from the user produced a new wrong answer instead of Claude actually reading the conversation line by line

Impact

This is a fundamental instruction-following failure. If Claude cannot execute "present this file in full" correctly — the simplest possible task — and then cannot accurately report what it did when asked, it undermines trust in Claude's ability to follow any instruction reliably. The flip-flopping under questioning is particularly damaging because it means Claude's self-reporting cannot be trusted.

Environment

• Claude Code CLI
• Model: Claude Opus 4.6 (1M context)
• CLAUDE.md with explicit "do exactly what the user says" as Rule 1

Additional Context

The user had to ask the same question 5+ times across 10+ conversation turns before Claude finally executed the instruction correctly. The conversation included explicit profanity-level frustration, and Claude still took multiple attempts to stop guessing and start following the literal instruction.