[BUG] Agent/Skill instructions are read but not followed proactively

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Summary

Claude Code reads instructions in CLAUDE.md and system prompts but does not proactively follow them. Explicit user instructions are required to trigger the expected behavior.

Examples

Example 1: Session startup instruction ignored

Setup: CLAUDE.md contains the following instruction:
At session start, read .ai/TODO.md and report any incomplete TODOs.

Expected: Claude reads TODO.md at session start and reports incomplete items.

Actual: Claude does not read or report until explicitly asked by the user.

Key point: When asked "Did you read the file?", Claude acknowledged that it had read CLAUDE.md, understood the instruction, but did not follow it.

Example 2: Agent not invoked despite matching task

Setup:

• System prompt states: "You should proactively use the Task tool with specialized agents when the task at hand matches the agent's description."
• issue-planner agent is defined with description: "Use this agent when creating work plans for GitHub Issues."

Task: User requests "Read issue #60 and create a work plan."

Expected: Claude invokes issue-planner agent via Task tool.

Actual: Claude creates the plan directly without invoking the agent, missing instructions defined in the agent (e.g., Test First methodology).

Impact

• Instructions placed in CLAUDE.md, skills, or agents may not be followed
• Users cannot rely on configuration files to enforce behavior
• Workaround: Users must explicitly instruct Claude to follow the rules or invoke specific agents

Environment

• Model: Claude Opus 4.5 (claude-opus-4-5-20251101)
• Tool: Claude Code Cli

What Should Happen?

Written in the lines preceded by 'Expected:' in the 'What's wrong?' section.

Error Messages/Logs

When asked about this behavior, Claude itself responded: 'Frankly, this is my limitation.' Claude acknowledged that it read the instructions but could not guarantee proactive compliance.

Steps to Reproduce

Written in the lines preceded by 'Setup:' in the 'What's wrong?' section.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.19 (Claude Code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

// settings.json
{
"permissions": {
"allow": [
"Bash(ls:*)",
"Bash(docker compose:*)",
"Bash(./gradlew:*)",
"Bash(direnv exec:*)",
"Bash(gh issue view:*)",
"Bash(gh issue list:*)",
"Bash(gh pr view:*)",
"Bash(gh pr list:*)",
"Bash(gh label list:*)",
"Bash(git status:*)",
"Bash(git diff:*)",
"Bash(git log:*)",
"Bash(git branch:*)",
"Bash(git check-ignore:*)",
"WebSearch"
],
"deny": [
"Bash(rm -rf:*)",
"Bash(git push --force:*)",
"Bash(gh repo delete:*)"
]
},
"hooks": {
"Stop": [
{
"matcher": "",
"hooks": [
{
"type": "command",
"command": ".claude/hooks/check-review-fix-queue.sh"
}
]
}
]
}
}