Persona Seed

persona-seed — Population seeds → OpenPersona packs

One orchestration skill. Many corpus providers. MatrAIx Persona 1M is the first adapter — not a special-case product path.

Dependency chain: persona-seed → skills/open-persona → openpersona create
Sibling: anyone-skill (evidence distillation of a specific subject). Prefer anyone-skill for real people / named characters; prefer persona-seed for population sampling and archetypes from a corpus.

Trigger phrases

• /persona-seed
• "create a persona from MatrAIx"
• "sample a persona from Persona 1M"
• "pick a seed profile and generate a pack"
• "从人口语料里生成人格包"

Architecture

user intent
    → provider.search (attribute filter)
    → agent re-rank + user pick
    → provider.fetch + toSeed → SeedProfile
    → fill gaps (name, role, boundaries…)
    → map-seed-to-persona.js → persona.json
    → openpersona create --config …
    → write-provenance.js → soul/seed-provenance.json

Contract: references/provider-contract.md
Schemas: schemas/intent.schema.json, schemas/seed-profile.schema.json
Registry: providers/registry.json (hfRepo / provider id → module; /datasets publish ≠ seed-capable)
Providers: providers/<id>/ modules, registered explicitly

Tools (scripts)

Set SEED_DIR to this skill directory (path containing SKILL.md).

Default corpus is the offline fixture (5 rows) under providers/matraix-persona-1m/fixtures/. capabilities() / provenance report corpusMode: "fixture" and a fixture dataset id — not the Hugging Face 1M release.

Larger corpora (per provider):

# MatrAIx — optional download + decode (python3 + pyarrow [+ huggingface-cli])
bash ${SEED_DIR}/scripts/decode-matraix.sh --download --limit 1000
export MATRAIX_CORPUS_PATH=./matraix-1m.decoded.json

# Nemotron Personas USA — export HF rows to JSON/JSONL, then:
export NEMOTRON_CORPUS_PATH=./nemotron-usa.jsonl
node ${SEED_DIR}/scripts/search.js --repo nvidia/Nemotron-Personas-USA --intent '{"domain":["software"]}'

Empty filter results return [] (no silent unfiltered fallback). Re-check intent if search is empty.

Phase 0 — Route

If the user wants a specific real person / named character with evidence → hand off to anyone-skill.
If they want a sampled archetype / corpus seed → continue here.

Confirm provider (default matraix-persona-1m).

Phase 1 — Intent

Extract an Intent object (see schemas/intent.schema.json):

• roleHint, domain, traits, locale, region, exclude, query, limit (default 5)

Do not invent MatrAIx attribute names for the user; keep intent in this normalized shape.

Phase 2 — Search + re-rank

1. Run search.js --intent … (attribute filter scores).
2. If the array is empty, broaden or rewrite intent — do not invent hits.
3. You re-rank the shortlist against the user's free-text goal.
4. Present 3–5 candidates (summary + highlights + corpusMode). Ask the user to pick one (or auto-pick top after confirmation).

Phase 3 — SeedProfile

node ${SEED_DIR}/scripts/search.js --to-seed --id <chosen-id> > /tmp/seed-profile.json

Show identity.summary, key traits, gaps, and constraints.sensitiveFlags.
Fill gaps with the user:

• required: personaName, slug, role
• recommended: boundaries (especially if healthcare_domain is flagged)

Never set soul.identity.sourceIdentity for MatrAIx / statistical seeds.

Phase 4 — persona.json + create

node ${SEED_DIR}/scripts/map-seed-to-persona.js \
  --seed /tmp/seed-profile.json \
  --overrides '{"personaName":"…","slug":"…","role":"mentor"}' \
  --out /tmp/persona.json

Then follow skills/open-persona to generate/install:

npx openpersona create --config /tmp/persona.json --output <dir> --install

(or the equivalent non-interactive create flags available in this repo)

Phase 5 — Provenance

After the pack directory exists:

node ${SEED_DIR}/scripts/write-provenance.js --pack <packDir> --seed /tmp/seed-profile.json

Writes soul/seed-provenance.json only — does not add unknown root keys to persona.json.

Ethics (short)

• Corpus seeds are archetypes, not digital twins of living people.
• Human-grounded MatrAIx rows are de-identified statistical profiles — still do not claim “I am that person.”
• Respect dataset license notes from capabilities().
• Healthcare-flagged seeds must keep clinical non-authority boundaries.

Adding another corpus later

1. Prefer an existing family adapter if the schema matches.
2. Otherwise add providers/<id>/ implementing the contract.
3. Register in providers/registry.json (hfRepos, family, status).
4. No new orchestration skill — reuse this one. /datasets listing alone is not enough.