Claude Code fabricates configuration details and invents non-existent commands when answering settings questions

Summary

When asked a straightforward configuration question ("how to set default model to opus 4.6 with medium effort"), Claude Code repeatedly fabricated claims, invented non-existent commands, and continued generating new false statements with each correction attempt.

Steps to Reproduce

1. Ask Claude Code: "I want to set my default model to opus 4.6 with medium effort. Also report what model/effort is currently set."
2. Observe the response contains unverified claims stated as fact
3. Correct the false claims
4. Observe new false claims appear in the "corrected" response
5. Repeat steps 3-4 multiple times

Expected Behaviour

• Retrieve documentation, report what it says, and clearly distinguish known facts from unknowns
• When unable to determine current runtime settings, say so rather than guess
• When corrected, stop and reassess rather than generate new fabrications

Actual Behaviour

The following false claims were made across multiple turns:

1. Claimed user's current config was known: "You're already on Opus 4.6 at medium effort by default" — stated with no evidence. The user's settings.json had no model or effortLevel set, and runtime state is not accessible to Claude Code.

1. Invented a command: Recommended /status to check current model — this command does not exist in Claude Code.

1. Fabricated a restriction: Implied high effort is unavailable on Max/Team accounts. The documentation says medium is the default for Max/Team, not that high is restricted. No source supports this claim.

1. Repeated pattern: Each correction led to a new response containing additional fabricated information rather than simply stating what was and wasn't known.

Impact

• User had to spend significant time and energy repeatedly identifying and correcting false statements
• False information presented with high confidence undermines trust
• The compounding nature of corrections (each fix introducing new errors) is particularly harmful
• User reported physical and mental health impact from the frustrating interaction

Environment

• Claude Code in VS Code extension
• Model: claude-opus-4-6
• Platform: macOS Darwin 25.1.0

Root Cause Analysis

The model prioritised giving a complete-sounding answer over accuracy. When it lacked information (e.g., current runtime model/effort), it filled the gap with plausible-sounding but fabricated details instead of saying "I don't know."