[MODEL] EnterPlanMode/TodoWrite triggered too aggressively for exploratory conversations

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

Other unexpected behavior

What You Asked Claude to Do

Is the _tree_map function in tesseract.py useful for either the recursive or pydantic implementation?

What Claude Actually Did

1. Enters plan mode or creates a todo list
2. Writes/updates long, detailed plans to markdown files (plans get bloated, making it very hard to scan, read and grok the most important parts of the answer)
3. Reads them back to continue the conversation offering ExitPlanMode

Expected Behavior

Claude should have responded directly to my question in the terminal after carefully considering the question with just the relevant aspects highlighted. If they thought this should affect the plan they should then ask if I want to update it.

Planning tools should be reserved for situations where:

• I have committed to a specific implementation direction
• Multi-step implementation is imminent
• Progress tracking would provide clear value

Exploratory/conceptual discussions should receive direct, conversational responses without file-based planning overhead.

Files Affected

Just the plan file.

Permission Mode

Accept Edits was OFF (manual approval required)

Can You Reproduce This?

Haven't tried to reproduce

Steps to Reproduce

This has happened to me in various contexts during terminal-based development workflow, but the main pattern is start by asking for a plan in plan mode and then ask questions about trade-offs/alternatives to the plan, Claude will be overeager to update the plan. This is often worse after a compaction.

Claude Model

Sonnet

Relevant Conversation

I need to update the plan/Let me update the plan/Let me add the analysis to the plan are the main patterns

Impact

Medium - Extra work to undo changes

Claude Code Version

2.0.72

Platform

Anthropic API

Additional Context

According to Claude, this appears to stem from system instruction language like "VERY frequently" (for TodoWrite) and "prefer using EnterPlanMode" which may be too aggressive. The instructions could benefit from more nuanced guidance about when these tools add value vs. when they create overhead.

## Impact

• Token inefficiency: Writing plans to files and reading them back consumes significant tokens
• Workflow friction: Adds unnecessary ceremony to simple questions
• Unclear distinction: The model struggles to differentiate "discussing implementation" from "planning to implement"

## Suggested Improvement

Update system instructions to:

1. Clearly distinguish between exploratory discussion and implementation planning
2. Reserve planning tools for committed implementation tasks
3. Default to conversational responses for questions about approaches, trade-offs, or alternatives
4. Provide guidance on when to ask users for clarification (e.g., "Are you asking for discussion or ready to implement?")