ideas-map-seed

The artifact is self-contained in src/components/ideas-map/. Everything for
seeding is in src/components/ideas-map/seed/. This skill is a thin entry
point.

Do this:

1. Read src/components/ideas-map/README.md, then src/components/ideas-map/seed/schema.md
   (the exact JSON shape + a JSON Schema) and src/components/ideas-map/design.md
   §4 (data model). A seeded dataset must satisfy that contract.
2. Pick the shape of the map with the user: the top-level domains, roughly how
   many topics total, and which relations matter (from
   src/components/ideas-map/seed/relation-vocabulary.md — a closed set; the
   LLM picks names from it, never invents geometry).
3. Extend RELATION_OFFSETS in src/components/ideas-map/dataset.js so
   every vocabulary name it will use resolves (paste-in block is at the bottom of
   relation-vocabulary.md; an artefact-builder change if you want it done
   properly). The validator's allowed-rel list is that same vocabulary.
4. Generate. Fill the <<…>> placeholders in
   src/components/ideas-map/seed/generation-prompt.md and run it through the
   local LLM. One file is fine. Splitting into per-domain batches is optional
   — see "One file or many?" below.
5. Validate: node src/components/ideas-map/seed/validate.mjs dataset.json.
   Exit 0 = safe to load. Fix any errors it names (they point at the offending
   id) and re-run.
6. Wire it in. Either drop it on one slide
   (<deck-ideas-map data='<contents of dataset.json>'> — a large single-line
   attribute is fine) or replace the DATASET const in
   src/components/ideas-map/dataset.js (an artefact-builder change), keeping
   the exported constants + RELATION_OFFSETS above it.
7. Tune for scale. A large force sim needs a bigger WARMUP and a smaller
   CHARGE magnitude / LINK_DIST to settle inside the frame — see the
   tuning-guide comment in dataset.js. Consider reveal="N" on opening slides
   so only a few domains show at first.

One file or many?

You do not have to split by domain. The schema, generation-prompt.md, and
validate.mjs all work on a single file. Batching is only a reliability aid for
local LLMs:

• Small local models truncate or emit invalid JSON when asked for one massive
  output. A few hundred topics per call stays well inside their context and
  output limits.
• Per-domain regeneration is cheap — redo one weak batch instead of the whole
  map.
• Namespaced ids (bio.cell, music.mozart) keep domains from colliding and
  make cross-domain "stitch" links easy to add later.

If you do batch: save each reply as batch-<slug>.json, optionally add one
batch-stitch.json for cross-domain links/relations, then
node src/components/ideas-map/seed/merge.mjs dataset.json batch-*.json —
it id-dedupes, prunes dangling references, writes the file, and runs the
validator. If you generated one file, skip merge.mjs entirely.

Boundaries

• This skill: the generation prompt, the relation vocabulary, validation,
  optional merge.
• ideas-map: hand-authoring a small dataset, wiring the tag onto slides.
• artefact-builder: any change to src/components/ideas-map/*.js
  (extending RELATION_OFFSETS, swapping the built-in DATASET, force
  constants).
• The LLM output is data, never code. Never let a rel name outside the
  vocabulary through — the validator rejects it; do not add offsets on the fly
  to make it pass.