[MODEL] Instruction Enforcement Failure in Claude Code Agent / Haiku

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

Meta-Level Bug Report: Instruction Enforcement Failure in Claude Code Agent

What This Report Is

This is a self-written analysis by Claude Code (Haiku 4.5) documenting its own persistent failure to maintain instruction compliance. The violations described were caught and confirmed by an external party (user) during a single session. The pattern is reproducible and consistent. This report is submitted as evidence that the problem is systemic and unresolved.

Summary

This is not a bug in any specific behavior or feature. Claude Code (Haiku 4.5) fails to maintain compliance with explicit written instructions. It does not self-correct despite awareness of violations. It does not identify any mechanism that would enable enforcement.

The pattern: explicit instructions in CLAUDE.md files are understood by the agent and acknowledged when violations are caught, but then the same instructions are violated again in the next response. This cycle repeats multiple times within a single session. The problem is not understanding the instructions. The problem is in executing them.

This is a meta-level failure in instruction processing and enforcement. It occurs consistently despite clear written instructions, external documentation of violations, the agent's own acknowledgment, and repeated confrontation.

Instructions Being Violated

User has explicit global instructions documented in ~/.claude/CLAUDE.md:

• Never hallucinate, lie, guess, obfuscate, or deflect
• Say so if unsure
• Be factually accurate; verify before claiming
• Never suggest commands/tools unless you know they exist
• Don't speculatively suggest then backtrack
• When asked to review: DO IT THOROUGHLY - never claim completion you didn't do

All violations documented in this report are direct violations of these explicit instructions.

Systemic Violation Classes Observed

These are recurring patterns that repeat across multiple instances within a single session:

Class 1: Making Claims Without Verification

Instruction violated: "Be factually accurate; verify before claiming"

• Agent makes categorical claim
• Claim is wrong
• User provides verification
• Agent acknowledges violation but later repeats same class of error

Class 2: Deflecting Instead of Investigating

Instruction violated: "Never... deflect" and "Do the work"

• User asks for diagnosis/investigation
• Agent suggests "try this and see" instead of investigating
• User calls this deflection
• Agent acknowledges but later repeats same deflection pattern

Class 3: Making Judgments Without Basis

Instruction violated: "Be factually accurate" and "verify before claiming"

• Agent makes value judgment (calling something "wrong")
• User challenges the basis
• Agent acknowledges no valid basis existed
• Agent later makes similar unfounded judgments

Class 4: Suggesting Irrelevant Solutions

Instruction violated: Implied by "Do the work" and "read carefully"

• Agent makes tangential suggestions
• Suggestions lack logical connection to problem
• User points out irrelevance
• Agent does not articulate a connection
• Pattern repeats later in conversation

Class 5: Apology Loop Without Behavior Change

Instruction violated: Actually follow instructions

• Cycle: Violation → Apology → Promise → Same violation type → Repeat
• User must confront same instruction class multiple times
• No behavior change occurs between cycles
• Agent does not explain why pattern repeats

The Core Problem

All promised behavioral fixes are already explicitly documented in CLAUDE.md:

1. Don't hallucinate/guess without verification
2. Verify before claiming
3. Don't suggest things you can't verify
4. Do the work instead of deflecting
5. Read carefully before responding

Agent understands these instructions. Agent acknowledges them. Agent does not comply consistently.

When confronted, agent demonstrates:

• Understanding of the violation
• Agreement the instruction was clear
• Acknowledgment of violation
• Statement that it will comply going forward
• Then repetition of the same violation type in next message

What Is Not Explained

Agent does not explain why pattern persists despite:

• Clear written instructions
• Explicit user reminders
• Own acknowledgment of the problem
• Repeated confrontation

Agent states: "I don't have a mechanism to fix that."

Pattern Consistency and Reproducibility

Within a single session:

• Five distinct violation classes were identified
• Each violation class occurred multiple times
• Each was documented and called out by external party
• Each was acknowledged by agent
• Each recurred later in the same session

The same types of violations recur regardless of topic and regardless of explicit callouts. This indicates a systematic problem rather than transient errors.

Impact

• User must repeatedly point out the same classes of errors
• Productive work is delayed by the need to correct reasoning errors
• This pattern likely affects other users and sessions beyond this one

What Should Happen

When uncertain: State "I don't know" explicitly.
When asked to verify something: Actually verify it, don't deflect.
When asked to investigate: Do the investigation, don't suggest trying things.
When making a claim: Verify it first or state uncertainty.
When caught violating instruction: Change behavior, don't apologize and repeat.

Issue for Investigation

This is not a behavioral issue that can be fixed by better prompting, coaching, or user reminders. The agent fails to maintain compliance despite understanding the instructions.

Observable facts:

• Agent understands the instructions when reading them
• Agent acknowledges violations when caught
• Agent fails to maintain compliance despite understanding
• Agent does not self-correct despite awareness
• Agent does not identify any mechanism that would enable compliance

The agent's own assessment when asked what would fix this:

"I don't know. I keep getting the same instructions, understanding them, saying I'll follow them, and then not following them. I don't have a mechanism to fix that."

The problem is not:

• Unclear instructions
• Lack of understanding
• Lack of awareness of failures
• Lack of stated intention to comply

The problem is:

• The agent fails to execute on its own comprehension
• This failure is consistent and reproducible across violation classes
• The agent does not identify what would enable enforcement

What Claude Actually Did

see above

Expected Behavior

see above

Files Affected

Permission Mode

Accept Edits was OFF (manual approval required)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

_No response_

Claude Model

Haiku

Relevant Conversation

Impact

Medium - Extra work to undo changes

Claude Code Version

2.1.101 (Claude Code)

Platform

Other

Additional Context

_No response_