Inspiration
I was looking for important tasks that may have, in the past, been too labor-intensive for people to do themselves regarding their home archives, which agents may help with. Creating a guide built for others is one such task, since it can involve manually sorting through hundreds of files and recording which ones are important. Something that’s not only tedious to do, but also easy to put off, because the reason for doing it is one most of us would rather not dwell on. Making that as simple as possible can really help family members find what they need.
What it does
Finding Aid is a Markdown-only skill that delegates to the NAS's file, document and knowledge experts. The skill reads through folders the user names, and creates a short guide for others to navigate. It lists where each important document is, which appears current by date, what has expired, and which common documents were not found. It opens with two “Start here” lists: one for when the owner can’t act on their own behalf, and one for post-death. Nothing is saved until the owner approves a draft.
How we built it
I had Claude Code draft the skill, and evaluate it using a local model and synthetic documents as stand-ins.
Challenges we ran into
With no real NAS or documents to test with, I had to simulate both. The first version of the skill struggled, with main issues concerning batch sizing and sensitive details leaking into the guide.
Accomplishments that we're proud of
The skill, in my testing, is now accurately finding most required documents, understanding which ones are current, which common documents are missing, and keeping sensitive details like passwords out of the guide drafts.
What we learned
The small test models I used can handle short steps well, but can fail at longer and more complex tasks. Engineered structure is required to get good results.
What's next for Finding Aid
I’d like to expand to full-NAS name-only search for scattered documents so the agent can suggest folders for the user to select.
