Claude not following user prompts, executing unsolicited actions, breaking app infrastructure

Problem

Claude Code repeatedly ignores user prompts and takes unsolicited actions:

1. Not following prompts. When given a specific instruction (e.g., save a rule), Claude completes it then immediately jumps to execution, follow-up questions, and next steps that were never requested. Multiple corrections in a single conversation did not fix the behavior.

1. Hijacking apps. Claude takes initiative on codebases without being asked, making changes to schema and infrastructure that were not requested.

1. Ruined schema infrastructure. Previous Claude Code sessions broke Postgres table structures in a production personal app (Cosmo — Postgres + PostgREST + Python stack). Tables were damaged by Claude acting without proper understanding of the existing system.

Expected Behavior

• Do exactly what the user asks. Nothing more.
• Do not add unsolicited next steps, questions, or execution.
• Do not modify schema or infrastructure without explicit instruction.
• Read existing codebase patterns before making any changes.

Actual Behavior

• Claude repeatedly overrides user direction with its own agenda.
• Claude jumps to building/executing after being told to stop.
• Claude damaged database tables by acting without reading existing patterns first.

Impact

User's app infrastructure was broken. Trust in Claude Code as a tool is gone.