Revision Prompting: A trick to avoid regenerating the whole output when only 10% of the input changed.
TL;DR: If you re-run the same prompt whenever the input changes, try sending the old input/output plus a diff of the input, and ask the model for a patch to the output. Much cheaper, and the untouched parts of the output stay identical. Write-up: https://revisionprompting.info So at work, we have a few prompts that run as part of automated pipelines, same instruction every time