Inspiration
A NAS is where footage goes to be forgotten. Editors, YouTubers and event photographers build up years of photos and clips, and then, when a new project starts, they cannot find the shot they know they already have. So they scrub through folders, reshoot it, or buy stock.
Most AI file tools start from a search box: you ask one question and get one list. But a creator does not start from a query. They start from a script or a shot list. I wanted a Skill that starts from that document and answers the question that matters before a shoot: what is already in the can, and what is still missing?
What it does
In the Can takes a script, a client shot list, or a voiceover recording, and checks it line by line against the photo and video library on a UGREEN AI NAS.
- It splits the brief into beats and shows them for confirmation. Abstract lines with nothing to see are flagged, not guessed at.
- It searches the local library for each beat, for stills and for clips.
- It checks the top still candidates with the on-device vision model.
- It returns a pull list: every beat marked Found, Thin or Missing, with real file paths and a confidence label.
- It returns a shoot list of everything still missing.
- On request, and only after platform approval, it copies the picks into a project bin, tags the originals, creates an album and saves both lists. Originals are never moved, deleted or overwritten.
- After the next import, Re-check mode searches only the missing beats among newly uploaded media.
The same workflow serves a video editor finding b-roll, and a wedding photographer checking a client's must-have list before delivery.
How we built it
The Skill is a single SKILL.md with five Markdown references, written to the official Uliya Skill Development Standard with the UGOS Skill Creation Tool (uliya-skill-kit 0.7.7, capability catalog 2.2.0).
It orchestrates four registered experts through standard delegate instructions:
- document-agent reads the script or shot list.
- knowledge-agent transcribes a voiceover and saves the finished lists.
- photo-agent runs content search, resolves named people, checks stills with the vision model and creates the album.
- file-agent reads clip details, copies picks into the bin and applies tags.
There are no scripts, no direct CLI calls and no external services. Everything runs on the local model, and nothing leaves the NAS.
I used Claude (via Claude Code) to research the capability catalog, draft the Skill and proposal, and run the official checker, and I directed and reviewed the work.
Challenges we ran into
- Video cannot be visually verified. The vision model reads still images only, so the Skill never labels a clip as verified. Clips are marked as a content match or a name match, and the pull list says so plainly.
- No code allowed. With no scripts, every rule has to be written so that a language model applies it the same way twice. I fixed the confidence labels and status rules in a reference table for that reason.
- One image per vision call. That makes checking slow at scale, so runs are capped at 20 beats and 30 visual checks, and the Skill reports which beats were left unchecked.
- Designing blind. Round 1 has no NAS access, and the catalog's sample outputs are synthetic. Every step is written to stop and report when a needed value is not returned, instead of guessing.
Accomplishments that we're proud of
- The package passes the official static checker with zero errors and zero warnings.
- "Missing" is treated as a real result, not a failure. It becomes the shoot list.
- The Skill is honest about its limits: every match carries a confidence label, and it never invents a file, a path or a person, or widens the search to force a hit.
- Every write action is copy-only and waits for platform approval.
What we learned
- The strongest case for an AI NAS is the archive that is too big to upload anywhere else. A footage library is exactly that.
- Reading the capability catalog closely matters more than the idea. Several promising ideas died because one step had no supported action.
- Small parameters open real features: filtering search by upload time is what makes Re-check mode possible.
What's next for In the Can
- Validate on a real UGREEN AI NAS with a licensed stock test set and a written answer key, including a brief where nothing matches.
- Find out whether the library indexes video by content, and tune the clip fallback to match.
- Tune the caps and the vision prompt against real speed and accuracy.
- Publish to the UGREEN Skills Hub.
Built With
- claude
- claude-code
- markdown
- ugos
- ugreen-ai-nas
- uliya
- uliya-skill-kit
