Inspiration
Creative teams revise briefs constantly — a new aspect ratio, a swapped shot, a renamed deliverable. The risk is mundane but expensive: sending a stale, same-name export from a private NAS to a client because nobody rechecked it against the latest brief. One editor described version-stack confusion pushed them to keep separate stacks per draft. We designed BriefProof for the team that has no integrated approval pipeline and still checks by hand.
What it does
BriefProof is an AI Skill that maps revised creative requirements to NAS delivery files for a pre-upload check. It confirms requirements and revision semantics, links each condition to its source, the deliverable, observed metadata, and SHA-256 — and keeps missing-file specifications visible instead of dropping them. Mismatches, unknowns, and human checks stay distinct: a landscape-to-portrait revision with unchanged story.mp4 bytes triggers review, because actual dimensions, not hashes, establish a mismatch. Rows cite observations and withhold "clear" when something is unknown. Reruns rehash against a confirmed baseline, and a ledger can be saved for the next round.
How we built it
SKILL.md orchestrates file, document, and knowledge experts through sub-agents (ug-filesystem, ug-filemgr, ug-document, optional ug-knowledge). Inputs are paths and confirmed requirements. Read-only by default, with admission limits (20 files, 10 GB) covering briefs and prior ledgers. No scripts, credentials, overwrites, or external services; unverified routing or canonical containment stops dependent processing, with no cloud fallback. Unpublished briefs, exports, and prior evidence stay together on the NAS, and local inference means selected footage is checked in place. The user retains approval.
Challenges we ran into
The hard part is honesty about uncertainty: model extraction and matching can err, missing metadata stays unknown, and hashes prove bytes, not correctness. NAS installation, routing, and expert behavior are still untested — local synthetic-media observations validate fixture facts, not Skill execution on a real NAS.
What's next
Validate with changed-brief fixtures: missing and mismatched exports, then same-name replacements. A full-replacement fixture has five deliverables, ten machine checks, and three human checks; both planted defects must be flagged, and a separate clean fixture must match all ten machine checks. Then measure latency and effort including confirmations, and test twelve edge cases including incomplete access.
Built With
- ai
- claude
- markdown
- nas
- sha-256
- ugos
- ugreen