knowledge-entities
Resolve and enrich Knowledge people, companies, projects, and concepts during an authorized ingestion or entity update, with sourced State, source notes, and links. Read-only identity questions do not authorize writes.Resolve and enrich entities
Find who or what an entity is before asking, decide whether it deserves a page, then write a sourced profile, its source note and links. Read a reference only when its step needs it. Shared references are identical across Knowledge skills; if another Knowledge skill brought you here, don't reread ones you already read.
- setup: always, first (status, ontology, search)
- identity: resolving an entity; whether to create a page
- enrichment: tiers, what to extract, external lookups, bulk runs
- profiles: page structure, creating or updating
- writing: before writing page prose
- sources: retaining sources and citing them
- assertions: claims and relationships about the entity
- publish: page writes and batches
- recovery: any uncertain, rejected or conflicting write
- verify: verification and report
- evidence repair: only for an authorized evidence-set repair
Contract
- Check Knowledge before asking the user or searching externally. Ask only with a hypothesis.
- No stubs or placeholders: every page's first line says who or what it is.
- Every fact has an inline citation. The source links to every enriched page, and each enriched page has a source note: its relationship with the source.
- Scale effort to importance. When in doubt, don't create a page.
- Never overwrite user-written assessments; flag contradictions with both citations.
Steps
1. Identify entities
List the people, companies, projects and concepts in the source. Separate actual participants from third parties mentioned in passing.
Done when: each entity is marked participant, substantive or incidental.
2. Resolve each one
Run the resolve chain in identity. Reuse an existing page's ID. Decide create, update or skip using its page-creation rules, and pick the best-fitting declared ontology type.
Done when: each entity has an existing ID, a create decision with a type, or is skipped; ambiguous identities are recorded as partial or unresolved.
3. Gather signal
Extract texture from the source and, for Key and Notable entities, look up what's missing (enrichment). Only supported claims move on; inferences stay labeled where they appear.
Done when: each entity has its supported facts and texture, each with a citation, or a reason it has none.
4. Write
- Create: introduction and State, plus only the optional sections the evidence supports (profiles). No Timeline section.
- Source note: one
attendedordiscussesrelationship between the entity and the source, with a one-sentence body (source notes). It is the entity's row in its derived Timeline. - Update: edit a State topic only when its meaning changed; use
knowledge_edit_bodyfor local edits. Retire a written Timeline section when the page still has one (retire a written Timeline). - Use declared ontology fields only; omit unknown values.
- Page bodies have a size budget, smaller for hub pages
(page budgets). When a write is
rejected, retire a written Timeline section first; otherwise give the subject
that outgrew its section its own page with
knowledge_extract_section. Never truncate. - Create claims and evidenced relationships for independently useful assertions (assertions). Markdown links don't create relationships.
Done when: every planned write has a published receipt.
5. Cross-reference
Link source → entity using observed paths. The way back from the entity is its source note, so the entity page needs no link to the source unless its prose cites it. Update related pages when the new signal matters to them: the company page from a person, and the reverse; the parent project from a sub-feature.
Done when: the source links to every entity and every entity has its source note.
6. Verify
Run verify with planned source → entity
expectedLinks, associations and source-note endpoints. Reread each changed page
against writing and the no-placeholder rule.
Done when: structural checks pass and every changed page reads as a real briefing.
Report
Return resolved IDs, created/updated/reused pages, sources used, unresolved or partial identities, contradictions and skipped work to the caller or user.