Oral History Time Capsule

Inspiration

Almost every family has a recording like this somewhere: a parent or grandparent telling how they met, the house they grew up in, the trip that changed everything. On a NAS, that recording sits a few folders away from thousands of scanned prints and old letters, but nothing connects them. The story is rarely replayed. When it is, nobody can quickly find the photo of "the blue scooter" or the invitation that proves the wedding was in 1986, not 1985.

I wanted the NAS to do what a patient family historian would do: listen to the story, find the pictures and papers it talks about, and gently point out where memory and records disagree.

What it does

You say one sentence, for example "Turn Grandpa's wedding story into an album". Oral History Time Capsule then:

1. finds the recording;
2. transcribes it, with speakers and timestamps;
3. extracts the people, places, years, events and objects that are mentioned;
4. finds supporting letters, invitations and certificates;
5. finds photos and videos that match, and confirms each one visually;
6. fact-checks every spoken claim as confirmed, corrected or unverified;
7. after your approval, creates an album and saves a memoir chapter. The chapter quotes the storyteller's own words and links every photo and document.

How I built it

It is a pure orchestration Skill, built with the UGOS Skill Creation Tool:

• SKILL.md plus two linked references, one for the matching rules and one for the memoir format;
• the parent Skill only delegates self-contained tasks to four registered experts (file-agent, knowledge-agent, document-agent, photo-agent) and passes forward only real results from earlier steps;
• steps 1–5 only read data;
• the only writes are one album and one new Markdown file, and each happens only after the user confirms;
• original files are never moved, renamed or deleted;
• everything runs on the NAS local models, so no family data leaves the house.

Challenges

• No code allowed. Every behaviour had to be expressed as delegation, approval and failure handling, so each step needed precise, self-contained instructions.
• Scanned photos lie about dates. Their capture date is the scanning date. Matching instead relies on named people, printed date stamps, folder names and a vision check of what is actually in the picture.
• Respecting memory. A corrected year must never silently overwrite Grandma's words, so the memoir shows both versions side by side.
• No device access in Round 1. I designed empty-result and partial-result paths for every step, so the Skill reports honestly instead of filling gaps.

What I learned

On an AI NAS, the value is not one clever model call. It comes from chaining trustworthy steps over private data, and from being explicit about what was verified and what was not.

What's next

On the real NAS I will validate three cases:

• a two-speaker recording with matching photos hidden among unrelated ones;
• a spoken year that contradicts a scanned invitation;
• a story with no matching photos at all.

After that, I plan to support multi-session stories that grow into a full family book.

Built With

• local-ai
• markdown
• ocr
• speech-to-text
• ug-document
• ug-filesystem
• ug-knowledge
• ug-photo
• ugos
• ugreen-ai-nas
• uliya-skill-kit
• vision-vlm