Rough Cut — Your NAS Library, Indexed by Mood

Brief it like an assistant editor and get a sequenced, mood-matched EDL from your movie, TV, and photo libraries — plus the themes your library actually supports.

Comment

Inspiration

Editors and creators keep terabytes of media on a NAS because that's where the archive lives. But a NAS indexes files; it doesn't understand editorial intent. Finding the twenty seconds of rain-on-glass you know you have means opening folders and scrubbing at midnight. It's the part of editing nobody bills for.

When we studied UGREEN's 30 official seed skills on the Skills Hub, one pattern jumped out. Every one retrieves or tidies: search photos, dedupe files, transcribe audio, organize folders. Not one produces an artifact the user wants. And in the Creativity track ("Turn your media library into a starting point for creativity"), no seeded skill does anything creative with media. That gap is the whole project: Rough Cut turns your NAS library into an edit.

What it does

You brief it the way you'd text an assistant editor:

"I need a 20-second opening with a lonely, rain-on-window mood."

Rough Cut sweeps your Theater (movie and TV) library and photo library, assembles mood-matched sequenced selects, and writes a CMX 3600-style EDL plus a readable shot manifest into your files (the format Premiere and Resolve open today). Alongside the sequence it reports two or three alternative themes your library supports, each backed by the matching-candidate counts behind it ("your library skews golden-hour coastal; here are three sequences it naturally tells"). It never invents a theme with zero results.

How we built it

Rough Cut is a Uliya UGOS Skill: a pure Markdown orchestration document that delegates all NAS work to registered experts, with no scripts or executables, exactly as the Uliya Skill Standard requires. One brief chains four capability domains no seed skill combines:

  1. Brief parse — an editorial mood vocabulary (8 mood classes with shot-grammar pacing data) maps the brief's words to controlled search selectors.
  2. Video sweep (media-agent) — ug-video search by style and keyword: the Theater library's own mood-aware search.
  3. Photo sweep (photo-agent) — semantic search-picture, then vision-vlm grades the top stills against the brief's mood descriptors (0–5 score, capped at 12 calls).
  4. Runtime fitting (file-agent) — mediainfo reads clip durations from the files so the sequence fits the target duration exactly.
  5. Composition — shots ordered into an arc (establish → develop → land) with proposed in/out timecodes and a one-line why-it-fits per entry.
  6. Alternatives — derived only from sweep hit counts; never invented.
  7. Artifact write — after an explicit confirmation gate, the EDL and manifest are written to the user's files.

Every delegation state is handled explicitly (completed / failed / cancelled / waiting_user / waiting_approval), and empty or partial results are reported accurately. The package passes UGREEN's own check-skill.mjs static checker with zero errors and zero warnings.

Challenges we ran into

The big one: our original design shipped ffmpeg scripts for keyframe extraction and scene sampling. The Starter Kit's hard rule ("Do not include scripts, executable files, or undisclosed external downloads") killed that plan just before build week. Rather than forcing the old design into the new box, we re-derived the concept from the capability catalog: the Theater library's style search and the photo library's vision model already cover perception; the missing piece was orchestration (brief → selects → sequence → EDL). The constraint made the skill more honest, not less capable.

The second challenge was scope. ug-video search covers the Theater library, not raw footage folders; we declared that in the skill itself instead of papering over it. "Your reference library, indexed by mood" is the product, stated plainly.

What we learned

The capability catalog is the product surface, and reading it end to end before we designed anything saved the project. The Uliya standard's honesty rules (never invent NAS data, declare unsupported steps) read like editorial discipline: an assistant editor who overclaims is worse than one who says "here's what we've got."

What's next

The engine (semantic search → cross-matching → sequencing over NAS media) generalizes. If we advance to finals, the same pattern powers Trace, a surveillance cross-reference investigation skill: sweep events, corroborate appearances across cameras, deliver an evidence-backed report. One engine, many verticals. Rough Cut itself goes live on the iDX6011 Pro, where Uliya's execution-chain UI shows every delegated step natively.

Built With

  • cmx-3600-edl
  • markdown
  • ug-filemgr
  • ug-filesystem
  • ug-photo
  • ug-video
  • ugos-skill
  • uliya
  • vision-vlm
Share this project:

Updates

Submission history