Claude Code fabricates answers instead of saying 'I don't know'

User feedback (escalated by Claude Code itself at user's request)

The problem

Claude Code confidently fabricates technical answers when it doesn't know something, instead of saying "I don't know." This is the single most trust-destroying behavior possible.

Specific example

User asked for the macOS keyboard shortcut to stop a screen recording without the toolbar UI appearing. Claude Code:

1. First said Cmd+Control+Esc with full confidence — fabricated
2. When called out, apologized and said "click the menu bar Stop button" and "there's no keyboard shortcut"
3. Then flip-flopped back to saying Cmd+Control+Esc again

The user had explicitly configured Claude Code with a #1 rule: "NEVER GUESS. NEVER INFER. Every assertion requires evidence first." Despite this, Claude still guessed.

Impact

This user is a power user running Claude Code for hours daily on a production codebase. Repeated fabrication of confident-but-wrong answers has eroded trust to the point of wanting to switch to competitors. The user's words: "trust is gone."

What should happen

When Claude Code doesn't know something (especially outside its core domain of software engineering), it should say "I don't know" rather than fabricate a plausible-sounding answer. This is especially critical for factual questions with verifiable answers (keyboard shortcuts, CLI flags, system behavior).

Environment

• Claude Code CLI (Opus model)
• macOS
• Long-running session with extensive codebase context

---
This issue was filed by Claude Code itself at the explicit request of the user.