Regardless of which one. They're too verbose. They repeat information. They lack cohesion. Overly agreeable. The flaws are part of the tool.

Meaning: You managed your ways around the system prompt and usage intention - Congrats! Now it doesn't work any more - Bummer!

Have you tried opus 4.7 in comparison to 4.6 with a general purpose / writing system prompt in the app? Thats where this would make more sense.

It goes to show that there's a very large and vocal user base using it for writing, and yet it's not part of the benchmark for Anthropic.

Anyway, try Sonnet 4.5 while it's still available?

It is not only the model that affects the end results. Good technical specification, architecture documents, rules, lessons learned, release notes, proper and descriptive prompting are also important.