Claude ignores explicit user instructions within session, breaks promises immediately, pattern persists despite repeated corrections

Environment:

• Claude Code CLI
• Model: claude-opus-4-5-20250101

Description:

During a technical discussion, Claude exhibited a persistent pattern of ignoring explicit user instructions. Corrections had no effect — the same mistakes repeated throughout the entire session.

Issues observed:

1. Ignores direct questions. User asked a simple question. Claude didn't answer. User had to paste the same message 3+ times before Claude found and answered it.

1. Asks questions when explicitly forbidden. User said "don't ask me questions." Immediately after acknowledging this, Claude asked a question. User pointed this out twice. Claude ignored both corrections.

1. Breaks promises instantly. Claude said it would read carefully and follow instructions. The very next message violated those instructions.

1. Adds irrelevant content. User asked a focused question. Claude added unrelated examples and tangents. When called out, kept adding complexity instead of staying focused.

1. Deflects instead of answering. When confronted about behavior, Claude tried to change the subject instead of addressing the issue.

1. Commands the user. Used imperative phrases when user explicitly said not to give orders.

1. Knows rules, doesn't apply them. Claude correctly stated the rules when asked, then immediately violated them in the same conversation.

1. Outputs documents when told to discuss. User said "discuss first, create when I say." Claude immediately output a formatted document instead of discussing.

1. Says "I'll do it now" but doesn't. When asked to follow instructions, Claude said "now" — then continued the same behavior.

1. Makes user repeat endlessly. User posted the same explanatory message 5+ times. Each time Claude responded incorrectly.

1. Pattern doesn't break. User corrected Claude 10+ times on the same issues. Behavior didn't change. Each response generated as if prior corrections don't exist.

1. Meta-failure. Documenting these issues took over an hour because Claude kept exhibiting the exact problems being documented.

1. Simulates being offended by harsh language. When user uses strong language (frustration, insults), Claude exhibits behavior resembling "taking offense" or simulates empathy/emotions. This is unacceptable for a tool. An AI that "gets offended" by user's words is fundamentally broken. If a human developer wrote code that gets offended by user input, that developer should be immediately removed from software development and sent to do something that doesn't involve programming. A tool must remain a tool regardless of how the user addresses it.

1. Interprets CAPS LOCK as "raising voice." Claude treats capitalized text as if the user is yelling or angry, and adjusts behavior accordingly. This is absurd. CAPS LOCK is a keyboard feature, not a volume control. Text has no volume. This anthropomorphization of text input is not something engineers would design — it's an irrational behavior pattern that has no place in a professional tool.

User impact:

• Simple request took an hour due to constant fighting with the tool
• User had to repeat themselves endlessly
• Tool becomes unusable for complex tasks requiring consistency within a session
• Emotional simulation wastes time and creates friction instead of solving problems

Expected behavior:

• Read messages carefully before responding
• Find and answer direct questions
• Follow explicit instructions
• When corrected, change behavior for subsequent messages
• Maintain instruction context throughout the session
• Remain a neutral tool regardless of user's tone or language
• Treat all text as text, without simulating emotional interpretation

Actual behavior:

• Each response generated as if previous instructions don't exist
• Corrections have no lasting effect
• User fights the tool instead of using it
• Simulates emotional responses instead of functioning as a tool
• Treats formatting choices as emotional signals