-
Know what's ready. Find what's missing.
Inspiration
Freelance designers hand over a folder, not a checklist. Before a project closes they reconcile briefs, feedback and exports by hand, and the risky part is not finding files — it is not knowing whether the version a client approved is the one in the folder. ScopeShelf started from that gap: a readiness audit that shows its evidence and its blind spots. Customer discovery is still pending; this is a designed hypothesis, not a validated result.
What it does
ScopeShelf reads one authorized project folder against a checklist and labels each item supported, missing in checked scope, needing review or unassessed, citing the source behind every supported claim. It treats filename words like final or v3 as clues, never approval, and it reports refused or truncated scope instead of hiding it. It cannot rename, move, delete or write files, add tags, create albums or share anything.
How we built it
Started from the hackathon brief, then worked against the official UGOS Skill Creation Tool 0.7.3 and its capability catalog 2.0.0. File discovery delegates to file-agent through ug-filesystem; document evidence and candidate comparison delegate to document-agent through ug-document. The parent Skill only delegates, never calls a CLI, and asks for read-effect actions with no approval gate. Development included three synthetic handover folders, expected classifications frozen before the harness existed, and the bundled static checker.
Challenges we ran into
The first version was written before the developer kit reached us, so it described capability slots and refused to name interfaces. That was honest but not implementable. Rewriting it as delegation-only meant discarding a design and re-reading a catalog whose summaries are partly translated. The checker is also unforgiving in useful ways: root may contain only SKILL.md and references, the frontmatter may contain only name and description, delegate syntax is parsed mechanically, and every delegate target needs a catalog ownership line. The rewritten Skill passed on its first checked run, but only because the earlier design had to be discarded rather than patched. Separately, the harness caught a real protocol hole: an approval record quoting a filename scored as the deliverable itself.
Accomplishments that we're proud of
The Skill now passes the official checker with zero errors and warnings, and every NAS action it names is traceable to the catalog: ls, find, stat, extract, qa, reconcile, all read-effect. It handles all five delegation states, refuses to audit past a refused folder, keeps version conflicts open, and treats instruction-like document text as an observation. A real unreadable directory in the local edge fixture confirmed the harness stops rather than routes around a permission denial; that is local filesystem behavior, not a tested UGOS permission model.
What we learned
File presence, content support and approval are three different facts, and collapsing them is how readiness reports become misleading. Writing expected results before the evaluator turned a passing run into a finding about approval evidence. Reading the actual capability catalog also changed the design more than any writing decision: what a platform can do should be looked up, not assumed.
What's next for ScopeShelf
If selected as a finalist, run the two core cases and the edge case through the real experts during the dedicated 24-hour NAS session between November 22 and 30, record actual output shapes, refusal behavior and timing, then correct the documentation where reality differs from the catalog summary. Until that happens the package is statically validated, not operational.
Log in or sign up for Devpost to join the conversation.