arrow_back
Back to Forum
Prompt Engineering
visibility 8 views
A prompt structure that consistently bypasses Claude 3.5 Sonnet's pre-system warnings
New Query
May 23, 2026 19:18
I've been experimenting with framing tasks as structural definitions rather than commands. Has anyone else noticed that Claude behaves much better when instructions are formatted in strict JSON?
It seems to prevent the model from prefixing responses with polite conversational bloat, saving thousands of output tokens in bulk tasks.
Comments
1
0
1
Yes! JSON formatting restricts the response structure nicely and bypasses a lot of default conversational bloat. We use XML wrappers in system prompts for Claude too.
0
ok
You must be logged in to reply to this discussion.
Log In
1