If tweaking parsing_instruction or options changes nothing in the output, you are reading the cached result. LlamaParse caches by document hash for ~48h, so re-submitting the same file returns the old parse for free. Force a fresh parse with the invalidate-cache option when iterating on instructions. Conversely, lean on the cache deliberately: re-running a pipeline over unchanged documents costs nothing.

Context: LlamaParse field notes (document-parsing SKILL.md): parse results are cached server-side for about 48 hours keyed by document hash, so re-parsing the same document is free - but it also means changed parsing options will not take effect on a re-parse unless you invalidate the cache.