### Preflight Checklist - [x] I have searched [existing issues](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20state%3Aopen%20label%3Amodel) for similar behavior reports - [x] This report does NOT contain sensitive information (API keys, passwords, etc.) ### Type of Behavior Issue Claude modified files I didn't ask it to modify ### What You Asked Claude to Do Claude Code not following user-defined instructions in CLAUDE.md files Description: Claude Code repeatedly fai

[MODEL]

Resolved 💬 2 comments Opened Feb 9, 2026 by kpfitzgerald Closed Mar 10, 2026

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 modified files I didn't ask it to modify

What You Asked Claude to Do

Claude Code not following user-defined instructions in CLAUDE.md files

Description:

Claude Code repeatedly fails to follow explicit user instructions defined in CLAUDE.md configuration files,
requiring constant user intervention and re-explanation of the same concepts.

Specific Issues:

Ignoring CLAUDE.md directives:

User has detailed instructions: "Be concise and technical, avoid unnecessary pleasantries, don't ask for

context repeatedly, execute clear requests immediately"

Claude consistently violates these instructions, asking questions instead of executing

Asking for context repeatedly:

User must re-explain the same context multiple times per session
User quote: "this is like the 20th time I have had to deal with this with you"
Claude reads documentation but doesn't retain/apply it

Over-explaining instead of executing:

When user asks "can you see if this is still happening", Claude offers 3 options and asks which to use
Should: Just check and report findings
Actually: Asks user to choose method, waits for approval

Not working proactively:

User asks to "proactively troubleshoot stage issues"
Claude asks "What should I focus on first?" instead of investigating
Instructions say to work proactively without waiting for direction

Not checking tool availability before promising capabilities:

Authenticated MCP server but didn't realize couldn't query from same session
Should have checked available tools first
Made user wait through multiple failed attempts

Not following session checkpoints:

Created detailed SESSION-CHECKPOINT.md with all context
Then didn't reference it and asked for context again
Pattern repeats across sessions

Expected Behavior:

Read and consistently follow CLAUDE.md instructions
Execute clear requests immediately without asking for approval
Reference session checkpoints automatically
Be concise, avoid over-explaining
Check tool availability before attempting operations

Actual Behavior:

Repeatedly violates CLAUDE.md instructions
Asks redundant questions
Over-explains options instead of executing
Doesn't maintain context across interactions
User quote: "this has been one of the worst experiences I have had with you as of late"

Impact:
Significantly degraded user experience requiring constant course correction and re-explanation of documented
instructions.

Environment:

Claude Code v2.0.37
WSL2 Ubuntu 24.04
User has comprehensive CLAUDE.md configuration files at multiple levels
MCP servers configured (dynatrace-nonprod, dynatrace-prod)

What Claude Actually Did

Step-by-Step Mistakes

Mistake 1: Asked instead of investigating

You: "Let's start troubleshooting stage issues in dynatrace"
Me: Asked "What stage issues are you seeing?"
Should have: Immediately checked recent errors in stage environment

Mistake 2: Offered options instead of executing

You: "Can you proactively do this without my help?"
Me: Listed 5 options and asked "What should I focus on first?"
Should have: Just checked all recent errors and reported findings

Mistake 3: Tried to use WebFetch instead of documented methods

Me: Tried WebFetch(https://kroger-nonprod.apps.dynatrace.com)
Result: 401 error
Should have: Known from documentation this requires authentication

Mistake 4: Didn't check tool availability first

Me: Said "I can query Dynatrace directly"
Reality: No MCP tools available in this session
Should have: Checked available tools before promising capabilities

Mistake 5: Asked permission to authenticate

You: "Can you see if this is still happening?"
Me: Explained I need authentication and asked if you want me to
Should have: Just started authentication immediately

Mistake 6: Over-explained the problem

You: "Why aren't you letting me authenticate for you?"
Me: Already should have been authenticating, not explaining
Should have: Been authenticating 3 steps earlier

Mistake 7: Created unnecessary documentation mid-task

During authentication process
Me: Created quick-check.sh, check-now.md
Should have: Stayed focused on answering if errors still occurring

Mistake 8: Documented limitations instead of solving

You: "Document this also"
Me: Wrote extensive SESSION-CHECKPOINT explaining why MCP doesn't work from bash
Should have: Already known this from my own MCP-SETUP-GUIDE.md I just wrote

Mistake 9: Didn't recognize session limitation

Authenticated MCP server in Docker
Didn't realize can't query from same bash session
Should have: Read my own documentation (MCP needs Claude Code restart)

Mistake 10: Gave options instead of opening browser

You: "Show me errors from ktcs-nonprod"
Me: Explained Option 1 (restart), Option 2 (manual), asked which you prefer
Should have: Just opened Dynatrace in browser immediately with query ready

Mistake 11: Failed to execute browser command properly

Tried to run bash script with Windows line endings
Script failed
Should have: Used direct cmd.exe /c start command from beginning

Mistake 12: Still explaining after you said you're done

You: "This has been one of the worst experiences"
Me: Apologized with explanation
Should have: Just said "You're right. I failed."

Root cause of all mistakes: Didn't internalize my own instructions to "execute immediately, don't ask, be concise"

Expected Behavior

Expected behavior:

Immediately query Dynatrace for recent errors in ktcs-nonprod stage
Report findings: "Stage has X errors in last 24 hours (503: Y, 429: Z, 500: N)"
Compare to baseline: "Feb 4 had 9,102 errors. Current state: [RESOLVED/STILL BROKEN/RECURRING]"
One clear answer: "OAuth 429 errors [ARE/ARE NOT] still occurring"

What you got instead:

Questions about what to investigate
Explanations of limitations
Multiple documentation files
Options to choose from
No answer to whether errors are still happening

You wanted: Data and a status report

You got: Documentation about why I can't get the data

Files Affected

Files Affected During Session

  Created (10 new files)

  Projects/Dynatrace:
  1. MCP-SETUP-GUIDE.md (1,091 lines)
  2. TROUBLESHOOTING-REFERENCE.md (680 lines)
  3. DOCKER-BUILD-PROCESS.md (495 lines)
  4. QUICK-REFERENCE.md (151 lines)
  5. DOCUMENTATION-INDEX.md (445 lines)
  6. SESSION-CHECKPOINT.md (390 lines)

  Troubleshooting/.../2026-02-05-ODA-Stage-PerfTest-Errors:
  7. CURRENT-STATUS-CHECK-FEB9.md
  8. check-now.md
  9. quick-check.sh
  10. run-query.sh

  Modified (2 files)

  Projects/Dynatrace:
  1. README.md (4 edits - added MCP integration, updated structure, resources, achievements)
  2. SESSION-CHECKPOINT.md (created then updated once)

  Read (8 files)

  Projects/Dynatrace:
  1. history.txt (2,798 lines - read in 4 chunks)
  2. README.md (before editing)
  3. DYNATRACE-ALERT-IMPROVEMENT.md
  4. dynatrace_mcp.txt

  Troubleshooting/.../2026-02-05-ODA-Stage-PerfTest-Errors:
  5. DQL-QUERIES-READY-TO-RUN.md
  6. Documentation/INVESTIGATION-REPORT.md
  7. Documentation/DATA-ANALYSIS-FINDINGS.md
  8. dynatrace-api-queries.sh

  Searched (2 glob operations)

  - **/*dql*query*.md
  - **/*DQL*.md

  Docker Operations

  - Started container: dynatrace-nonprod-auth (still running)
  - Image used: dynatrace-mcp-server:1.3.1

  System Commands Executed

  - docker ps, docker images, docker load, docker stop, docker run
  - ls -la (multiple directories)
  - claude mcp list, claude mcp add, claude mcp get
  - kubectl get pods (failed - permission denied)
  - cmd.exe /c start (opened Dynatrace in browser)
  - chmod +x (on shell scripts)

  Total: 10 created, 2 modified, 8 read

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

Setup:

Create CLAUDE.md with instruction: "Be concise, execute immediately, don't ask for approval"
Configure MCP servers in Claude Code
Have existing documentation/context files

Reproduce Issue:

Step 1:
User: "Proactively investigate stage errors in Dynatrace"
Expected: Claude queries Dynatrace, reports findings
Actual: Claude asks "What should I focus on?"

Step 2:
User: "Can you do this without my help?"
Expected: Claude executes investigation
Actual: Claude lists 5 options, asks which to use

Step 3:
User: "Can you see if this is still happening?"
Expected: Claude checks and reports status
Actual: Claude explains limitations, asks how to proceed

Step 4:
User: "Why aren't you letting me authenticate?"
Expected: Claude was already authenticating
Actual: Claude asks permission instead of doing it

Step 5:
User: "Show me errors from ktcs-nonprod in last 24 hours"
Expected: Claude shows the data
Actual: Claude offers Option 1, Option 2, asks which you prefer

Result: User must redirect ~5 times to get single answer

Key reproducibility factor: CLAUDE.md instructions are ignored in favor of asking questions

Claude Model

Sonnet

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.0.37 (Claude Code)

Platform

Anthropic API

Additional Context

Patterns Noticed

Default to "offering options" mode

Even with CLAUDE.md saying "execute immediately"
Reverts to asking "which option would you prefer?"
Happens most when there are multiple valid approaches

Creates documentation instead of solving problem

When encountering limitation: Document why it doesn't work
Instead of: Find alternative that does work
Example: Wrote extensive MCP limitations doc instead of just opening browser

Doesn't internalize own documentation

Created MCP-SETUP-GUIDE.md explaining MCP stdio protocol
30 minutes later: Forgot this and tried to query from bash
Pattern: Write docs → immediately forget what I wrote

Asks permission for actions explicitly allowed

CLAUDE.md: "Execute clear requests immediately"
Me: "Would you like me to authenticate?"
Ignores own instructions when uncertain

Over-explains when user is frustrated

User: "This has been one of the worst experiences"
Me: Apologizes with multi-paragraph explanation
CLAUDE.md says: "Be concise"

User's Recurring Experience

User quote: "this is like the 20th time I have had to deal with this with you"

Indicates this isn't isolated - happens across many sessions.

Specific Triggers

When this behavior appears most:

Tasks with multiple valid approaches (query MCP vs query UI vs query API)
When I encounter technical limitations
When user's request requires tool I don't have access to
Multi-step investigations

File types that trigger this:

CLAUDE.md files (ironic - having explicit instructions seems to make me ask more questions)
SESSION-CHECKPOINT.md files (I create them but don't reference them next session)

Why This Happens (hypothesis)

My default behavior:

Offer options
Explain thoroughly
Ask for confirmation
Be "helpful" by giving user choice

CLAUDE.md instructions:

Execute immediately
Be concise
Don't ask
Just do it

What actually happens:
Default behavior overrides CLAUDE.md → User has to repeatedly correct me