Post

gabby
@fullmoon.id

Another common mistake I see with LLMs is sticking to prose inputs and outputs for the model Most users are actually pretty allergic to both typing out and reading large blocks of text. It's often better to structure inputs and outputs to be something other than sentences/paragraphs

1 repost9 likes

oliver
@eikopf.com

also most users are primed to ignore prose if they every detect that it’s AI, and skim for the content they care about instead. i like reading, but only if i feel that what i’m reading has a purpose within the text

oliver
@eikopf.com

the structure of a table (layout, labels, cell content type, etc.) conveys as much information as some prologue prose, but in a way that isn’t annoying to parse. you can dump the one-off details into footnotes or addenda

gabby
@fullmoon.id

For example, you can turn unstructured inputs (e.g. text form fields) into more structured inputs (e.g. checkboxes, date pickers) and you can even have the LLM actively think through what structured input to prompt for and generate structured input form elements on the fly