omd-scout

name

omd-scout

description

Build a measured reference board without designing anything: whole pages for feel, tight selectors for component anatomy, typography studies, motion studies, image refs for the unrenderable. Use when the user asks for references, inspiration, benchmarks, or "how do good sites do X" — standalone, before or without a build. Minimum 18 captures, target 25. Triggers: 레퍼런스 찾아줘, 레퍼런스 수집, 참고 사이트, 벤치마킹, find references, inspiration board, how do other sites do.

omd-scout

A reference board, measured instead of pinned. This skill only collects and reasons — it designs nothing.

Spawn the omd-scout subagent with the concept (ask one short question if none exists and no .omd/frame.md provides one) and the thing being studied. If the brief contains URLs, pass them to the subagent — they are captured first, marked as user-provided, and exempt from the famous-site quota. The subagent starts with a component inventory — every nav, card, form field, and footer the build will need, listed before the first capture — then works to a floor of eighteen captures, targeting twenty-five:

omd ref add <user-url> --as <name> --from-user                   # user-provided URL — captured first, quota-exempt
omd ref add <url> --as <name>                                    # whole page — rhythm and feel (3–4)
omd ref add <url> --as <name> --selector ".nav"                  # one component's anatomy (one per inventory item)
omd ref add <url> --as <name> --selector ".nav" --blueprint      # full-resolution structural blueprint — max 3 per board
omd ref add <url> --as type-study-1                              # chosen for its typography — minimum 2
omd ref add <url> --as type-study-2                              # different register, different pairing
omd ref add <url> --as motion-study-1                            # chosen for its motion — minimum 2, minimum 4 if brief mentions animation or register is showpiece
omd ref add <url> --as motion-study-2                            # different domain, different vocabulary
omd ref add <pin-url> --as mood --image                          # unrenderable — reasoning only
omd ref add <reddit-url> --as list-debate --image                # community: what people felt and why (minimum 2)
omd ref add <url> --as voice-study --image                       # how a real product writes
omd ref principles <url> --as <name> --add "..."                 # why it works, one sentence
omd ref list

Blueprints: use sparingly, max 3. A blueprint (--blueprint --selector) captures the full node tree of one component with the skin abstracted to color roles. Use it only when the component solves the exact structural problem the build faces, or when the user pointed at a specific component. A blueprint is not pixels — it is measurement at full resolution bounded to one component, with color roles replacing literal values. Structure transplants; skin never does; the page-distance guard still fires on a clone.

Before the first capture the scout runs at least six WebSearch queries — problem domain, craft, community, typography, motion, and competitor — and logs them all on the board. Famous sites (Linear, Stripe, Vercel, Notion, Raycast) may not exceed one third of the total captures; they earn their place or they stay off the board.

When the register is showpiece, the board must also draw from award galleries — these are "who solved a problem of this intensity and won doing it" sources, not just "who built this well" sources. Mandatory showpiece sources: awwwards.com (SOTD archive), godly.website, thefwa.com, siteinspire.com, minimal.gallery. For Korean briefs, add gdweb.co.kr. Each award-gallery capture must include a one-sentence answer to "what made this site win" — stated as the effect on the viewer, not the name of the technique used. Award gallery captures count toward the floor only when they carry this sentence plus the full principle set.

Every capture gets a one-line justification, principles that answer why, a note on what contradicts the concept, and one line naming what specific measurements from this capture are candidates for the build — which token group they would feed (type scale, motion durations, spacing ladder, radius ladder) and why. This line is the upstream source for .omd/attribution.md: the hand traces each token decision back to a named capture; without this line, the attribution table has no origin to cite. omd ref add prints a design-signal score; a page under 0.4 does not count toward the floor.

Rejection obligates replacement. When a capture is rejected — for slop (2+ findings), for low signal, for kinship, for a blocked-page error — it does not leave a hole. The board is made of what passed, not of what was available. Every rejected capture triggers a replacement search: a different query from a different angle, a source tier at least one step higher in the trust hierarchy. The scout does not stop until every slot in the composition contract is filled with a capture that passed. Anti-references — captures kept to document what to avoid — do not count toward the floor: their value is in the warning they carry, not the slot they occupy.

When omd ref add throws a blocked-page error, the URL is not retried. Demote to --image if you can construct honest reasoning from public discourse about the site; if not, discard and run the replacement search.

Never describe how a reference looks (Jansson & Smith: shown an example, designers reproduce its features even after the flaws are flagged; a model is worse). Numbers and reasons only. The board lands in .omd/refs/ and any later omd-ultradesign run picks it up automatically.