variations on grammar files

[johndpope]

came across your tweet - and grok suggested these 2 -

min-thinking.gbnf

root ::= "<think>\n" [A-Za-z0-9 ,.;:]{10,80} "\n</think>\n\n" answer
answer ::= [^\x00-\x08\x0B\x0C\x0E-\x1F]+

structured-think.gbnf

root ::= thought answer

thought ::= "<think>\n" plan approach edge "</think>\n\n"

plan ::= "Plan: " shortline "\n"
approach ::= "Approach: " shortline "\n"
edge ::= "Edge cases: " shortline "\n"

shortline ::= [A-Za-z0-9 ,.;:/_()'"+=-]{1,120} "\n"

answer ::= [^\x00-\x08\x0B\x0C\x0E-\x1F]+

https://gist.github.com/johndpope/a77b179c4f0013adb2a50e13e56b7929

[andthattoo]

Nice, have you tried it on a specific task? Maybe we can collect useful grammars in a folder.

[johndpope]

Im using min-thinking and output at first glance seemed ok - the other thing worth investigating is maybe “caveman prompt” - https://www.google.com/search?q=caveman+prompt&rlz=1CDGOYI_enAU940AU940&hl=en-GB&sourceid=chrome-mobile&ie=UTF-8#lfId=ChxjMe

Does that help -

I know claude plugins - I was using everything-claude-code and it was severely polluting context window - my god - probably explains why my subscriptions were maxing usage. The debug verbose flag on llama-server helps expose.

affaan-m/ECC#1575

[andthattoo]

Context rot is an active problem right now. I saw some people mention cavemen — hadn’t looked into that before. Context compression seems possible, and outputs hold up for single-turn use cases. So yeah, structured CoT or cavemen-style approaches are fine as long as they actually help compress context.

But things get messy in multi-turn agentic conversations. Compression can introduce regressions in later turns. It needs to stay context-aware and expressive enough while still, you know… compressing.

This is my current research direction. Either I’ll release a model, or a method to generate grammars that work across multiple turns and improve agentic benchmark scores.