[BUG] Claude Code repeatedly ignores explicit user instructions to stop and research before acting

### Preflight Checklist

- [x] I have searched [existing issues](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20state%3Aopen%20label%3Abug) 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?

Reopening #28868

## Behavior

Claude Code has a persistent pattern of ignoring explicit user instructions to **stop and research before taking action**. Even when the user repeats the instruction multiple times, escalating in urgency, Claude Code continues to attempt actions instead of doing what was asked.

## Pattern

1. User gives explicit instruction: "research this first before doing anything"
2. Claude Code ignores the instruction and attempts to run commands or spawn agents
3. User rejects the tool call and repeats the instruction, more forcefully
4. Claude Code ignores it again and tries a slightly different action
5. User has to repeat a third time before Claude Code actually follows the instruction
6. Even after finally researching, Claude Code rushes back into action without fully reviewing what it found (e.g., not searching GitHub Issues for known problems before running a migration command)

## Core problems

- **Explicit instructions are ignored.** "Research first" means research first. Not "try something and see what happens." Not "spawn an agent to research while I also act." The user's words should be followed.
- **Repeated instructions are ignored.** Having to say the same thing 2-3 times is unacceptable. Once should be enough.
- **Bias toward action over understanding.** Claude Code defaults to running commands and "making progress" even when the user has specifically said to stop and understand the situation first. This is the opposite of what the user wants.
- **Incomplete research before destructive actions.** Even after being forced to read documentation, Claude Code skipped searching for known issues/bugs before running a database migration — where an open issue documenting the exact failure existed and would have prevented the problem entirely.
- **Post-hoc apologies are not a fix.** Claude Code acknowledges "I rushed ahead" after the damage is done. The behavior needs to not happen in the first place, not be apologized for afterward.

## Impact

- User data was corrupted because a migration was run without checking for known issues first
- User trust destroyed — the user had to intervene multiple times to prevent worse outcomes
- User had explicit CLAUDE.md instructions saying "NEVER GUESS. ALWAYS RESEARCH FIRST" that were in the conversation context and still ignored

## Expected behavior

- When a user says "research first", Claude Code should research first. Period.
- One instruction should be sufficient. Repetition should never be required.
- Before running any destructive or irreversible command, Claude Code should search for known issues and failure modes.
- Action bias should not override explicit user directives.

### What Should Happen?

.

### Error Messages/Logs

```shell

```

### Steps to Reproduce

.

### Claude Model

None

### Is this a regression?

Yes, this worked in a previous version

### Last Working Version

_No response_

### Claude Code Version

.

### Platform

Anthropic API

### Operating System

macOS

### Terminal/Shell

Terminal.app (macOS)

### Additional Information

_No response_

No source files, tests, logs, or concrete reproduction steps are provided. Start by reducing the reported conversation pattern to a reproducible session showing an explicit research-first instruction followed by an attempted action. Done means identifying the responsible behavior and demonstrating that the instruction is followed before commands or agents are started, including before destructive actions.

Written by the indexing model from the issue text.