[BUG] incorrect analyze

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?

Title: Noticeable drop in task-following accuracy starting ~2026-06-15

Summary
Over the past few months I've used Claude Code for complex, multi-step engineering work (a Telegram/Bale/Rubika/Instagram sales-bot admin panel with scripting, segmentation, rules engines, commission logic, etc.) and it consistently handled deep, intricate tasks well — often going beyond what I explicitly asked for in a good way.

Starting yesterday (2026-06-15), I noticed a clear regression in how well it follows specific instructions during multi-turn work.

Specific examples

1. I described a "brain script" (bot conversation script) feature and attached a screenshot of the relevant admin panel. Instead of working on what I showed/described, it went and modified an unrelated part of the codebase.
2. Separately, I asked for changes to a rules-editing feature involving a two-way relationship with the system (rule conditions tied back to system state/fields). Again, it went off and did something unrelated instead of addressing what I asked.

What I'd expect

• Stick to the specific file/feature/screenshot referenced in the request rather than substituting a different task.
• If something is ambiguous, ask a clarifying question rather than guessing and editing the wrong area.
• Maintain the same level of precision on complex, context-heavy tasks that it previously demonstrated.

Impact
This makes me less confident handing off non-trivial tasks without closely double-checking every change, which reduces the efficiency gains I was previously getting.

What Should Happen?

Claude should stay focused on the specific file/feature/screenshot referenced in the request instead of substituting an unrelated task. If the request is ambiguous, it should ask a clarifying question rather than guessing and editing the wrong area. It should maintain the same precision on complex, context-heavy tasks that it previously demonstrated consistently.

Error Messages/Logs

Steps to Reproduce

1. Open Claude Code in an existing complex codebase (admin panel with multiple feature modules).
2. Describe a specific feature to work on (e.g. "the bot conversation script feature") and attach a screenshot of the relevant panel/screen for context.
3. Observe that Claude edits an unrelated part of the codebase instead of the feature described/shown.
4. In a separate request, ask for a specific change to a rules-editing feature (e.g. rule conditions tied to system fields/state).
5. Observe that Claude again works on an unrelated area instead of addressing the actual request.

This behavior started consistently around 2026-06-15, after previously handling similarly complex multi-step requests accurately.

Claude Model

None

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

last

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_