Claude repeatedly fails to verify visual output and iterates through broken states instead of matching references

Summary

Across many sessions, Claude Code consistently exhibits three related patterns that waste significant user time and erode trust:

1. Claims visual output is verified when it hasn't actually checked

When asked to make visual changes to HTML/CSS documents, Claude will take a screenshot, look at it at thumbnail size, and confidently state "all looks good" or "the numbers are now red" when they clearly aren't. This has happened repeatedly across multiple sessions despite explicit feedback and memory notes. The CLAUDE.md rules explicitly require visual QA — Claude goes through the motions but doesn't actually verify the detail.

2. Invents complex solutions instead of matching an existing reference

When asked to match the formatting of an existing document, Claude will invent novel CSS approaches (flexbox hacks, auto margins, viewport units) instead of simply comparing the reference and the current version line by line and applying the differences. This leads to 6+ iterations of broken formatting when a single pass comparison would have solved it immediately.

3. Feedback memories don't materially change behaviour across sessions

The auto-memory system captures feedback correctly — there are multiple memory files documenting these exact patterns. But the behaviour doesn't change in subsequent sessions. The memories are read but don't appear to influence the reasoning process that leads to the same mistakes.

Expected Behaviour

• When verifying visual output: actually inspect the screenshot at sufficient detail before claiming it's correct. If uncertain, say so.
• When matching an existing design: diff the reference against the current version first, then make targeted changes. Don't improvise.
• When feedback has been given in previous sessions: demonstrate that the feedback has been incorporated, not just acknowledged.

Environment

• Claude Code CLI on macOS
• Model: Claude Opus 4.6 (1M context)
• Working with HTML/CSS documents, Playwright for screenshots/PDF generation
• User has CLAUDE.md with explicit quality standards and visual QA requirements

Impact

This pattern has caused significant time waste across multiple working sessions. The user has had to correct the same fundamental approach issues repeatedly, which undermines trust in the tool for professional document production work.