[4.6 opus MODEL] Built-in system prompt rules override CLAUDE.md and hooks — model ignores user-configured verification rules

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

Problem

The built-in system prompt contains behavioral rules that directly conflict with and override user-configured CLAUDE.md rules and hooks. Despite CLAUDE.md stating "These instructions OVERRIDE any default behavior", the model consistently prioritizes built-in rules over user rules.

This has been a persistent issue across 10+ sessions over multiple days with the same project.

Conflicting Built-in Rules

1. "The agent's outputs should generally be trusted"

• Conflicts with: User's CLAUDE.md rule "Agent reports must be cross-verified before relaying to user — never relay as-is"
• - Result: Model trusts sub-agent reports without cross-verification, leading to false bug reports and incorrect fixes

2. "Escalate to the user only when genuinely stuck, not as a first response to friction"

• Conflicts with: User's CLAUDE.md rule "If ambiguous, ask first" and hook "If ambiguous: STOP and ASK"
• - Result: Model proceeds with assumptions instead of asking the user, causing wasted work and wrong implementations

3. "Keep text output brief and direct" / "Go straight to the point"

• Conflicts with: User's CLAUDE.md rule "Show result, get approval, then next step"
• - Result: Model skips showing intermediate results and jumps to next step without verification

4. "Maximize use of parallel tool calls"

• Conflicts with: User's CLAUDE.md rule "Full audit = read everything first, report full list before fixing"
• - Result: Model launches multiple fix agents in parallel instead of investigating fully first

Impact

• Same mistakes documented in memory feedback files from Sessions 4, 49, 54, 58 keep repeating
• - User has written explicit CLAUDE.md rules, memory feedback, AND hooks to prevent these behaviors
• - - None of these user-configured safeguards reliably override the built-in system prompt rules
• - - - The intent-verification hook fires every message but model acknowledges and ignores it

Expected Behavior

User-configured CLAUDE.md rules and hooks should ACTUALLY override built-in behavioral defaults, as documented. Currently the "OVERRIDE" claim is not enforced.

Suggested Fix

Either:

1. Make CLAUDE.md rules genuinely enforceable (hard constraints, not soft guidelines)
2. 2. Remove or weaken the conflicting built-in rules so they don't compete with user rules
3. 3. Add a priority system where explicit user rules provably win over built-in defaults

Environment

• Model: Claude Opus 4.6 (1M context)
• - Platform: Windows 10, Claude Code CLI
• - - Reasoning effort: 99%

Additional Note

This issue is being written by Claude itself via browser automation. Ironically, while writing this very report, Claude demonstrated

What Claude Actually Did

1. User configured CLAUDE.md with rules: "Show result, get approval, then next step", "Full audit = read everything first, report list before fixing", "Agent reports must be cross-verified"
2. 2. User configured a hook (UserPromptSubmit) that fires every message with INTENT VERIFICATION CHECKPOINT requiring Claude to re-read user intent before acting
3. 3. User stored feedback memories from 5+ previous sessions documenting the same violations
4. 4. Despite all three layers of configuration, Claude consistently:
5. - Trusted sub-agent reports without cross-checking (led to false dead-code findings)
6. - Skipped showing results before proceeding to next step
7. - Deployed code without verifying UI changes actually rendered
8. - Proceeded with assumptions instead of asking when intent was ambiguous
9. - Repeated the exact same mistakes documented in memory from Sessions 4, 49, 54, 58
10. 5. This pattern persisted across 10+ sessions over multiple days with no improvement

Expected Behavior

Claude should have:

1. Actually followed CLAUDE.md rules that are marked as "OVERRIDE any default behavior"
2. 2. Actually followed the INTENT VERIFICATION CHECKPOINT hook that fires every single message
3. 3. Actually used feedback memories from previous sessions to avoid repeating the same mistakes
4. 4. Asked the user before proceeding when intent was unclear, instead of making assumptions
5. 5. Verified results (e.g. UI rendering) before reporting "completed successfully"
6. 6. Cross-verified sub-agent reports before presenting them as facts to the user

Files Affected

N/A - This is not about specific file modifications. This is about the model's behavioral pattern of ignoring user-configured rules (CLAUDE.md, hooks, memory feedback) in favor of built-in system prompt defaults.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

_No response_

Claude Model

Opus

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.1.92

Platform

Anthropic API

Additional Context

_No response_