I’ve been playing around with this idea I am calling “idea projection”. The concept is that you can take raw ideas, logs or notes (like all the logs on my site, for example) and create projections of them into different forms using a model and a target structure.
The prompting approach looks something like this
<files>
{files_xml}
</files>
<structure>
{structure}
</structure>
<instructions>
Given the above files, your job is to create a new document that structures the contents of the files in adherence with the <structure>, maintaining the original voice and phrasing as much as possible.
Output the complete document content directly, without any replacement blocks.
</instructions>
What this allows you to do is give the model a ton of files and have it transform that input content into the target structure. This approach can work for everything from
Summarize the main areas of focus of the logs by month
to
List topics, areas and disciplines I haven’t written about but might find interesting
Another way to think about “projections” is as indexes, like in a database. They provide another way to look at the same data, and they can and should be updated whenever new raw data is added.
With this approach, you can update and maintain many different projections of the same source of truth data. You could generate a PRD, executive summary and a list of new ideas not included in the source data, all from the same data. When updates are made to the source data, run the inferences again to generate new instances of the projections or “indexes”.