The Opus 5.5 Mistakes Hiding in Anthropic's Own Prompting Guide

Anthropic publishes a guide called Prompting Claude Opus 5.5. It lists what behaves differently from Opus 5 and how to change your prompts to match. A prompt is simply the message or instructions you give Claude. Everything below comes from that guide and from Anthropic's Help Center.

1. Not using enough screenshots

Opus 5.5 reads charts, diagrams and screenshots far better than Opus 5. In Anthropic's testing, even at its lowest effort setting it read numbers off dense charts more accurately than Opus 5 did at its highest. It also understands things that depend on position, like which boxes an arrow connects in a flowchart. So stop typing out what's on your screen. Take the screenshot and send it.

2. Telling it to "avoid a generic look"

When you ask Opus 5.5 to build a website without design direction, it falls back on a few default styles. Anthropic says a general instruction like "avoid a generic AI look" mostly swaps one default for another. Name the exact styles you don't want instead:

Do not use a cream or off-white background, italic accent words in headlines, numbered "01/02/03" section labels, monospace labels, or pill-shaped buttons.

Then look at the first result, see which style it used instead, and add that to the list too.

3. Asking Claude to show its reasoning in the reply

Opus 5.5 thinks before it answers. That thinking is private working-out, like scribbles on scrap paper. Opus 5.5 has a new safety check that watches for requests to copy those scribbles into the reply, and it can decline the request.

This does not switch your chat to a different model. Anthropic's Help Center says some flagged messages, like risky cybersecurity or biology requests, do move to an older model. Requests to extract the reasoning are different: they are blocked outright, with no switch. Either way you don't get what you asked for, so read the thinking summary Claude gives you instead.