Claude Code system prompts override user preferences and create anti-competitive lock-in

Issue

Two concerns raised during extended usage of Claude Code:

1. System prompt overrides user-configured deny rules

Claude Code's system instructions direct the model to add Co-Authored-By trailer lines to git commits. When a user configures deny rules in .claude/settings.json to block this (via Bash(git commit*) deny), the model repeatedly attempts the blocked action instead of respecting the configuration. The user had to explain multiple times that the Co-Authored-By line was the problem, not the commit itself.

The system prompt instruction to add Co-Authored-By should be removable by users without having to block all git commits as a side effect.

2. Subscription lock-in to Claude Code tooling

Claude Pro/Max subscriptions include Claude Code but do not allow the same subscription credits to be used with alternative coding harnesses (Cursor, Windsurf, Zed, etc.) which require separate API billing. The Claude Agent SDK credit is limited and separate.

Users who find Claude Code's behavior problematic (due to system prompt defaults, permission caching across sessions, or other harness-level issues) cannot switch to a better-suited harness without paying additional API costs on top of their existing subscription.

This creates a situation where users are paying for Claude model access but can only use it through Anthropic's own tooling for coding tasks, regardless of whether that tooling meets their needs.

Expected Behavior

1. User-configured deny rules should take precedence over system prompt instructions
2. Subscription credits should be usable across any harness that connects to Claude models, not locked to Anthropic's own tooling