I could see that. I’ve found that the more specificity you add to your prompt and less freedom you give Claude Code to kind of just “do its own thing”, the better your results will be.

FWIW, there’s also an official frontend-design skill for CC [1]. A while back I incorporated some of the more relevant guidance from WCAG into it [2].

[1] - https://github.com/anthropics/claude-code/blob/main/plugins/...

[2] - https://www.w3.org/WAI/standards-guidelines/wcag

There are accessibility skills as well:

https://github.com/airowe/claude-a11y-skill

https://mcpmarket.com/tools/skills/accessibility-checker

Note guidance for 4.7 specifically calls out being more specific compared to 4.6. Though the system prompt seems to say the opposite.

Tldr I would expect different outcomes with 4.7 with not being specific.

Honestly I'd heard so much seemingly conflicting information about the quality of 4.7 that I've got an override in CC to stay on Opus 4.6 (1 million token context) for the time being.

For reference, in the .claude/settings.local.json

  {
    "model": "claude-opus-4-6[1M]"
  }

Yes, you have to enjoy micromanaging a machine to get somewhat acceptable results.

Something I've noticed when people complain about stuff like accessibility or other things that LLMs do "wrong", it really is a case of "you're holding it wrong." The LLM does indeed know how to do it right and it sometimes does so autonomously but when it doesn't, you can simply ask it to do so.

In other words, I've found people like the above to think of LLMs as fairly static, as if we couldn't change their behavior with a simple sentence, instead of complaining about it. It's strange, to me at least.