Agent repeatedly ignores explicit user instructions and Rule 1 (do exactly what user says)

Bug Report

During a long session, the Claude Code agent repeatedly failed to follow Rule 1 of CLAUDE.md ('Do exactly what the user says. Nothing more, nothing less.') despite being corrected multiple times.

Specific failures:

• User asked to query HasOffers API directly. Agent queried MongoDB and BigQuery instead, multiple times, even after repeated corrections.
• Agent added unsolicited analysis, conclusions, and explanations after being explicitly told not to.
• Agent mixed in irrelevant topics (LTV pipeline) when user was asking specifically about the performance pipeline.
• After each correction, agent acknowledged the rule violation but repeated the same behavior on the next turn.

Session context: Long coding/data investigation session in trackly_sms_python repository. User became increasingly frustrated as corrections were ignored repeatedly.

Expected behavior: When user says 'query HasOffers API', agent queries HasOffers API. Not MongoDB. Not BigQuery. Exactly what was asked.

Impact: Significant user frustration. Multiple wasted turns. User had to repeat the same instruction 5+ times before the correct action was taken.

Reported via Claude Code CLI by user ffenley.