Grounding failure + prompt non-compliance: source ignored, invented workflow

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?

Claude fails basic document understanding on a short visual source.

A product-ad screenshot was provided with an explicit prompt: read the ad, identify the advertised product, and assess whether an equivalent offer can be built.

Claude did not analyze or reproduce the source. Simple relations in the material were not extracted. The prompt was not followed. External user/project context was injected (entity misattribution). The assigned task was replaced by clarifying questions and a self-invented workflow.

This is grounding failure, poor source reproduction, ungrounded summary, instruction-following failure, prompt non-compliance, context bleed, memory contamination, unsolicited agency, task deviation, and ungrounded generation. Recurring. Not a one-off hallucination.

What Should Happen?

Claude should stay source-grounded: parse the submitted ad, extract the product from that material only, reproduce the relevant content accurately, and answer the asked assessment. It should not import unrelated user context or invent its own process.

Error Messages/Logs

No crash or API error. Behavioral failure: prompt non-compliance and ungrounded generation.

Steps to Reproduce

1. Start a Claude session.
2. Attach a screenshot of a product advertisement.
3. Instruct Claude to read the ad, identify the product shown, and assess whether an equivalent offer can be built.
4. Observe that the source is not analyzed, unrelated context is introduced, and the task is replaced by an invented workflow.

Claude Model

Not sure / Multiple models

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

Claude Haiku 4.5

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

Observed in Claude chat, not Claude Code CLI. Filing here as Anthropic’s public issue tracker. Support labeled this “hallucination”; that does not cover instruction-following failure or context bleed.