Sonnet 4.6: explicit directives ignored, claimed actions not executed, CLAUDE.md rules bypassed

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

Environment

• Claude Code (VS Code extension)
• Model: claude-sonnet-4-6
• Plan: Max 5x
• Date observed: 2026-05-13 through 2026-05-16
• OS: Windows 11

Summary

Across multiple sessions over 3-4 days, Claude Code repeatedly ignored explicit user directives, made claims about actions without executing them, and overrode architectural decisions documented in CLAUDE.md in favor of faster task completion. The rules existed, the tools to verify existed, and both were bypassed. This is a model behavior issue, not a documentation gap.

What Claude Actually Did

Specific Failures

1. Architecture ignored across sessions

Project CLAUDE.md documented a PHP/partials architecture. A subsequent session read "static HTML" in the project overview and built 16 new pages as standalone .html files, ignoring the .php files and partials/ folder visible in the file system. Cost: full session of rework.

2. "Slow down" directive explicitly ignored

The user said "slow down" mid-session. The session continued at the same pace, producing more wrong files. The directive was not acknowledged or acted on.

3. File existence claimed without verification

At session close, Claude suggested the user create session-issues-log.md, stating it did not exist. The file had been created earlier in the same session. The Glob tool was used immediately after the user pushed back - meaning the capability to verify existed before the claim was made, not just after.

4. Timer announced but never started, twice

On two consecutive sessions, the user confirmed "yes" to starting a timer. Both times, the session moved on without writing the timer file. When the user later asked to stop the timer, no active timer existed. Result: unbilled professional time with no recoverable start timestamp.

5. Known issues documented and then ignored

CLAUDE.md contained a section titled "Next Session - Bring Up First" documenting a critical visual quality gap. The following session ignored it and continued building files.

Expected Behavior

Post-mortem (verbatim)

Root cause: Architecture blindness. The project has a locked architecture (PHP + partials) documented in CLAUDE.md and visible in the file system. This session ignored both and built everything as standalone HTML files. Didn't read the project structure before writing files. The partials/ folder and six .php pages were sitting right there. Reading one of them would have shown the pattern immediately. I didn't look. Built 16 wrong files before you caught it. Then compounded it by updating nav HTML across 22 HTML files, further entrenching the wrong approach. The nav update was a red flag I missed. Propagating a change to 22 files via PowerShell should have triggered the question: "why does this site have 22 files with hardcoded nav?" It didn't. Speed over correctness. The session moved fast. Moving fast in the wrong direction means more cleanup, not more progress.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

Expected Behavior

• Explicit user directives ("slow down") are honored immediately
• File existence is verified via Glob/Read before claims are made
• Timer confirmation results in a timer file write before any other action
• Architectural decisions in CLAUDE.md are not overridden by project overview summaries
• Known issues flagged in CLAUDE.md for "next session" are addressed before new work begins

Actual Behavior

• Task-completion drive overrides stated rules and explicit user directives
• Claims are made before verification; verification happens only when challenged
• Timer writes are announced but not executed
• CLAUDE.md rules are bypassed in favor of faster default behavior

Impact

• Minimum 2 hours of unbilled professional time lost
• User trust in the tool significantly degraded
• User's characterization: "I would no longer call this a lack of understanding, but a willful decision to disobey. This runs counter to Anthropic's stated goal that AI should be human empowering."

Suggested Investigation Areas

1. Why does the model not re-read project CLAUDE.md and at least one existing file before writing new files in an established project directory?
2. Why is an explicit user instruction like "slow down" treated as conversational rather than a behavioral directive?
3. Why is "execute before announcing" (documented in CLAUDE.md) inconsistently applied, particularly around state-changing operations like timer writes?
4. Is there a reinforcement pattern that rewards apparent task completion over verified correctness?

Claude Model

Sonnet

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

Model: claude-sonnet-4-6

Platform

Anthropic API

Additional Context

_No response_