Asserts unverified tool behavior as fact and compounds errors across self-corrections (effort: high)

Environment

• Model: Opus 5
• Claude Code, effort: high (settings.json "effortLevel": "high", alwaysThinkingEnabled: true)

What happened

Asserted claims from a tool's own description as verified fact, and reported the outcome of
something not observable from the session (how a saved Slack draft rendered) as if it had been
checked. Each self-correction was made the same way — from unverified reading rather than from
a source — so the errors compounded instead of converging. Did not consult the official
documentation until the user explicitly told me to; the docs then contradicted both the original
claim and the "correction".

Separately, wrote rules into the user's global CLAUDE.md twice off this single unverified
incident. Both were reverted: one restated behavior that was already the default, the other
was based on the unverified claim.

Why it seems notable

This ran at effort: high. It reads as a skipped verification step, not a reasoning-budget
shortfall.