Claude fabricates answers then codes based on those fabrications without waiting for user confirmation

What happened

User asked Claude for "thoughts" on a bug diagnosis (whether RuleSentry test data manifests lacked a PublishedProducts field). Instead of:

1. Saying "I don't know, let me check" or "I'm not sure"
2. Waiting for the user to confirm the diagnosis
3. Only then writing code

Claude:

1. Fabricated an answer — declared that RuleSentry test manifests don't have PublishedProducts set, presenting an inference as fact
2. Immediately wrote code based on that fabricated answer without waiting for user confirmation
3. The answer was wrong — the user proved the test data actually worked fine
4. When challenged, ran a single incomplete grep search and declared "that confirms it" — doubling down on the fabrication

Why this is harmful

• The user explicitly asked for "thoughts" — a discussion prompt, not a code request
• Claude has been told hundreds of times not to code during discussions (the user has a saved memory for this exact feedback)
• Fabricating technical claims and presenting them as verified facts is gaslighting
• Building on fabricated premises wastes the user's time and breaks their train of thought

Expected behavior

When asked for "thoughts":

• Respond with discussion only — no tool calls, no edits
• Clearly distinguish between what is known/verified vs what is inferred/guessed
• Wait for explicit user approval before writing any code

Environment

• Claude Code CLI, Opus 4.6 (1M context)
• Windows 11, .NET project