Output verbosity

[mcljot]

A trend I'm seeing is verbose output, where the LLM tries to solve too many problems, gives too many discrete examples, etc. Need to tune/refine this and constrain the volume of output to what's practical and useful.

[rohennes]

A related point for the record:

Verbosity
There were a number of modules produced by docs-workflow such as feature benefits, feature limitations and requirements, feature examples, feature how-it-works, when a lot of these could be combined either into a single intro, concept topic, or added to the short description/prereqs for procedures.
The feature itself was a relatively small improvement to be able to add custom names to network interface cards. But the output of docs-workflow meant it would take up significant territory in the related assembly if all of the modules were included, overinflating the importance of the feature. So the challenge is how to stop over-documenting and recognize the relevant importance of feature updates.
https://github.com/openshift/openshift-docs/pull/109421/changes#diff-0fba04c2a7772803be48aaa827a20f9f46be482060ded05e5092d6b223038589

[aireilly]

I wonder are we overloading the context at various steps and silently failing on skills etc.

[aireilly]

if skill prompts are being dropped altogther the model would fall back to default slop mode.