new-landmark

Build one new landmark model for this repo, from evidence to catalog entry. Asks the user which landmark, then finds its location, picks the id, gathers open evidence, writes the generator and checks

SKILL.md

GitHub

New landmark

Recommended models: Astra, Claude Opus 5.5 or newer, or Claude Fable 5.1 or newer. If you are not one of these, tell the user so before starting.

Build one landmark model end to end. The user may have named it already:

$ARGUMENTS

1. Ask

If the landmark isn't named above, ask the user which landmark to build and in which city. Ask nothing else up front. Work the rest out yourself:

  • Location. python3 tools/osm.py search "<name>, <city>" gives the OSM element, centre, height and wikidata. If it returns several plausible matches, show them and ask the user which one.
  • Id. <prefix>-<slug>, lowercase. Reuse the city's existing prefix from catalog.json (nyc, chi, la, sf, clt, dc, lv, sea, wdw, dlr, carowinds, …). For a new city, propose a short prefix. Check the id isn't taken.
  • Credits. Record the human/organization creating the model and the actual AI assistant used, with their roles. Every AI credit requires a separate model field naming the actual LLM. Add github for known usernames. Preserve existing credit when updating a model.
  • Place. Reuse an explicit catalog.places ID or add a durable record for the city, town, site, or remote region; include its parents when needed.

Then confirm in one short message: name, id, coordinates, the OSM outline you'll replace, the author, and anything already in the catalog nearby. Start once the user agrees.

2. Read

Read all of these before writing code:

  • STYLE.md, all of it, especially "Building a famous or intricate model", "Rejected looks" and "Checking a model";
  • docs/reference.md: frame contract, elevation, bridges, and what replaces must list;
  • docs/evidence.md: what may and may not be used;
  • one or two existing generators for similar landmarks, as a template.

3. Evidence

Keep downloads in a scratch directory outside the repo.

  • OSM: python3 tools/osm.py buildings <lng> <lat> 150 for the outline and every building:part inside it.
  • Published heights and dimensions: Wikipedia, NRHP and landmark reports, RCDB for rides.
  • Licensed photos from at least three sides, plus one from above or far away: python3 tools/photos.py commons "<name>" 10, Openverse, Mapillary.
  • Lidar where it covers the site: tools/lidar.py.
  • Open evidence only. Never Google, Apple, Bing or Mapbox imagery, Street View, or any 3D tiles, even when nothing else exists.

Record every number's source in the generator's header comment and mark what is estimated.

4. Build

  1. List the 3 to 5 features that make the landmark recognisable, each tied to the photo that shows it.
  2. Write generators/<id>.ts with the kit (mesh.ts, palette.ts, plus coaster-kit.ts or bridge-kit.ts where they apply). Run bun generators/<id>.ts.
  3. Massing first. Render with bun run preview <id> preview/<id> <photo.png ...> and compare each render with a photo from the same direction. Fix proportions until the silhouette matches before adding any detail.
  4. Add the identifying features, then the facade. Check the 200 and 80 px views: if a feature vanishes, make it bigger or bolder, not more detailed. Stay within the budget in STYLE.md.
  5. Night. Decide what is lit on the real landmark after dark (signs, crown, floodlit stone, windows, lamps) and give each its night material from STYLE.md, "Night": lit signs neon* in their own colour, glass roofs glass. Check bun run preview <id> --night beside a night photo.
  6. Optional: check the footprint. Plot the OSM outline and parts from step 3 to scale, with north up, and compare them with preview/<id>/top.png. Walls, setbacks and corners should land on the OSM lines, within a metre or two, unless photos or lidar show OSM is wrong. Worth doing for large or irregular plans: stadiums, terminals, campuses, anything that isn't a simple tower.

5. Catalog and checks

  • Add the models entry with file, source, collection: "landmarks", type, tags, license (default CC0-1.0), and structured credits; every AI credit includes model. Record visual references and attribution as needed.
  • Add the landmarks entry with explicit place, anchor, bearing, scale, optional Wikidata, and every covered OSM outline/part in replaces.
  • For updates, preserve model/placement IDs and existing metadata, and regenerate both Low and High LOD files when present. Git supplies revision history.
  • Refresh stored building outlines with bun tools/replacement-footprints.ts when replaces changes.
  • bun run validate, bun run typecheck, bun test, and bun run build must pass. Inspect the built site for LODs, history, credits/avatars, all metadata, and day/night rendering. See docs/website.md.

6. Report

Show the user the three-quarter render beside a photo, and the phone-size views. Name every difference you can still see between the renders and the photos; never say it looks right without putting a render next to a photo. List the sources, the estimates and any doubts. Ask whether to iterate or to commit and open a pull request with a contact sheet (bun run sheet <id>).

Public catalog metadata

Use the metadata contract in docs/reference.md. Add structured credits for the actual human/organization and AI assistant, including its required model. Add a primary type, descriptive tags, and collection: "landmarks". Choose an explicit existing place for the placement, or add a durable place record for its city, town, site, or remote region. Do not infer authorship or place from defaults. Updating a model keeps its existing IDs and file paths so Git-backed website history remains intact.