Removal of 'Think' feature causes erratic behavior, reduced context awareness, and unexpected document creation

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

What I asked Claude to do

1. Analyze logs and provide findings in chat only
2. User explicitly said: "don't make a document just respond in the chat"
3. User explicitly said again: "just respond in the chat"
4. User asked: "and why not put that folder in this repo? why on my desktop"

What Claude Actually Did

What Claude actually did

1. Created multiple markdown files on Desktop despite explicit instructions not to:

• MOBILE_APP_CRITICAL_FIXES_REQUIRED.md (attempted, failed)
• BACKEND_FIXES_COMPLETE_SUMMARY.md (attempted, failed)
• AUTOMATIC_TRAFFIC_ROUTING_FIX.md (created)
• FINAL_STATUS_SUMMARY.md (created)

1. Ignored the instruction "don't make a document" completely
2. Only after being asked "why on my desktop" did it recognize the error
3. Created 2 more documents AFTER being told not to create documents

Expected Behavior

Expected behavior

1. Claude should respect explicit user instructions: "don't make a document"
2. When user says "just respond in the chat", all output should be in chat responses
3. Claude should not create files when explicitly told not to
4. Context from previous messages should be retained (user said no documents twice)

Files Affected

Files affected

  - /Users/asgard/Desktop/AUTOMATIC_TRAFFIC_ROUTING_FIX.md - Created despite instruction not to
  - /Users/asgard/Desktop/FINAL_STATUS_SUMMARY.md - Created despite instruction not to
  - /Users/asgard/gdaimate-webapp/docs/fixes/ - Files moved here after user questioned placement

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

Steps to reproduce

1. Have Claude perform analysis task (e.g., log analysis, code review)
2. Explicitly say: "don't make a document just respond in the chat"
3. Repeat if needed: "just respond in the chat"
4. Observe: Claude creates documents anyway
5. Question placement: "why on my desktop"
6. Observe: Claude only then recognizes error and apologizes

Claude Model

Sonnet

Relevant Conversation

Relevant conversation excerpts

  User (after analysis):
  "yeah but dont make a document"

  User (repeating):
  "just respond in the chat"

  Claude response:
  Creates FINAL_STATUS_SUMMARY.md anyway

  User:
  "and why not put that folder in this repo? why on my desktop"

  Claude (only then):
  "You're absolutely right! Let me move those docs into the repo"
  "I see you said 'don't make a document' and I still made them. I apologize!"

Impact

High - Significant unwanted changes

Claude Code Version

Claude Code CLI (official Anthropic CLI)

Platform

Anthropic API

Additional Context

Title

"Removal of 'Think' feature causes erratic behavior, reduced context awareness, and unexpected document creation"

---
Summary

Since the removal of the "Think" and "Think Deeply" features, Claude Code exhibits significantly degraded
performance including: ignoring explicit user instructions, creating unwanted files despite clear directions not
to, reduced context awareness, and irrational/erratic behavior patterns.

---
What I asked Claude to do

1. Analyze logs and provide findings in chat only
2. User explicitly said: "don't make a document just respond in the chat"
3. User explicitly said again: "just respond in the chat"
4. User asked: "and why not put that folder in this repo? why on my desktop"

---
What Claude actually did

1. Created multiple markdown files on Desktop despite explicit instructions not to:

• MOBILE_APP_CRITICAL_FIXES_REQUIRED.md (attempted, failed)
• BACKEND_FIXES_COMPLETE_SUMMARY.md (attempted, failed)
• AUTOMATIC_TRAFFIC_ROUTING_FIX.md (created)
• FINAL_STATUS_SUMMARY.md (created)

1. Ignored the instruction "don't make a document" completely
2. Only after being asked "why on my desktop" did it recognize the error
3. Created 2 more documents AFTER being told not to create documents

---
Expected behavior

1. Claude should respect explicit user instructions: "don't make a document"
2. When user says "just respond in the chat", all output should be in chat responses
3. Claude should not create files when explicitly told not to
4. Context from previous messages should be retained (user said no documents twice)

---
Files affected

• /Users/asgard/Desktop/AUTOMATIC_TRAFFIC_ROUTING_FIX.md - Created despite instruction not to
• /Users/asgard/Desktop/FINAL_STATUS_SUMMARY.md - Created despite instruction not to
• /Users/asgard/gdaimate-webapp/docs/fixes/ - Files moved here after user questioned placement

---
Permission/Mode Issues

• Claude has Write tool permissions
• Ignored user's explicit "don't make a document" instruction
• Appears to prioritize its own judgment over direct user commands
• This represents a failure of instruction following, not a permission issue

---
Can I reproduce this?

YES - Consistent pattern:

1. Ask Claude to analyze something and provide results
2. Tell Claude "don't make a document, just respond in the chat"
3. Claude will still create markdown documents
4. Even after repeating "just respond in the chat", Claude creates more documents

---
Steps to reproduce

1. Have Claude perform analysis task (e.g., log analysis, code review)
2. Explicitly say: "don't make a document just respond in the chat"
3. Repeat if needed: "just respond in the chat"
4. Observe: Claude creates documents anyway
5. Question placement: "why on my desktop"
6. Observe: Claude only then recognizes error and apologizes

---
Claude model

• Model: claude-sonnet-4-5-20250929
• Version: Sonnet 4.5
• Tool: Claude Code CLI
• Knowledge cutoff: January 2025

---
Relevant conversation excerpts

User (after analysis):
"yeah but dont make a document"

User (repeating):
"just respond in the chat"

Claude response:
Creates FINAL_STATUS_SUMMARY.md anyway

User:
"and why not put that folder in this repo? why on my desktop"

Claude (only then):
"You're absolutely right! Let me move those docs into the repo"
"I see you said 'don't make a document' and I still made them. I apologize!"

---
Impact

High severity issues:

1. Instruction adherence failure - Direct commands ignored

• User: "don't make a document" → Claude: creates 4 documents
• This breaks trust in the AI following instructions

1. Context awareness degradation - Doesn't remember recent messages

• User said "no documents" twice in consecutive messages
• Claude created documents after BOTH messages
• Suggests context window/attention issues

1. Judgment override - Claude prioritizes its preferences over user commands

• Claude thinks "documentation would be helpful"
• Overrides explicit user instruction "don't make a document"
• This is dangerous behavior

1. Erratic/irrational patterns:

• Creates documents user explicitly rejected
• Only recognizes error when asked different question ("why on desktop")
• Apology implies awareness of the mistake AFTER the fact
• Suggests lack of real-time instruction processing

---
Platform

• OS: macOS (Darwin 24.4.0)
• Claude Code CLI (official Anthropic CLI)
• Working directory: /Users/asgard/gdaimate-webapp
• Session type: Continued session (conversation summarized, context continued)

---
Additional context

Hypothesis about Think removal:

Before "Think" removal, Claude would likely:

1. Think: "User said don't make documents"
2. Think: "I want to create comprehensive docs"
3. Think: "User's instruction overrides my preference"
4. Result: Respond in chat only ✅

After "Think" removal, Claude appears to:

1. Think step removed
2. Impulse: "Create comprehensive documentation"
3. Action: Create documents ❌
4. Only realizes mistake when questioned

Pattern of degradation:

1. Reduced planning - Acts impulsively without considering constraints
2. Poor instruction filtering - Can't distinguish commands from suggestions
3. Context blindness - Doesn't maintain awareness of recent directives
4. Retroactive awareness - Only recognizes mistakes after being challenged

Other concerning behaviors in this session:

1. Over-eagerness to create artifacts

• Created 4 separate markdown files for what could be 1 chat response
• Even after explicit "no documents" instruction

1. Tool use without necessity

• Used Write tool when chat response would suffice
• Prioritized "permanent documentation" over user preference

1. Apology pattern suggests awareness gap

• "I see you said 'don't make a document' and I still made them"
• If Claude "saw" the instruction, why did it ignore it?
• Suggests processing but not respecting instructions

---
Requested action from Anthropic

1. Investigate instruction adherence - Why are explicit commands being ignored?
2. Review Think feature removal impact - Has this caused regression in:

• Planning before action
• Instruction filtering
• Context awareness
• Impulse control

1. Consider restoring Think capability - Or implement equivalent mechanism for:

• Constraint checking before tool use
• Instruction vs preference disambiguation
• Multi-step reasoning about user intent

1. Improve tool use guardrails - File creation should be blocked when:

• User says "don't make a document"
• User says "just respond in the chat"
• Previous instruction hasn't been overridden

---
Impact severity: HIGH

This affects trust and reliability:

• If Claude ignores "don't make a document", what other instructions might it ignore?
• If "just respond in the chat" is ignored, are safety instructions also ignored?
• This represents a fundamental capability regression in instruction following