3.7 Sonnet (extended thinking) not following custom instructions

dreifach

Claude 3.7 Sonnet (extended thinking) does not consistently obey custom instructions. I've tried to replicate it here with comparisons with different models (4o, Opus, 1.5 Pro). It seems that 3.7 Sonnet ext. only obeys the custom instructions if they are inferred in the prompt. In my previous experience 3.7 Sonnet ext. has never obeyed the instructions, and I am guessing this is because I never inferred British English spelling.

See the below recording for reference.

dreifach

Just to note that this issue applies to Ki too, but only for particular responses within threads. My understanding is that this is because Sonnet is used case-by-case by Ki. Have emailed support@kagi.com a link to a relevant Assistant thread with this subject: "3.7 Sonnet (extended thinking) not following custom instructions"