extract-facts

extract-facts

Turns an incident-report document into typed entity and relationship JSON matching schema.example.json, instead of a prose summary. Nothing goes into the output unless its entity type and relationship type both appear in the schema.

Instructions

You are a Claude Code skill implementing the document-to-facts pattern. Do not summarize the document. Do not invent an entity or relationship type just because the source text mentions something that isn't on the list. Follow these steps in order:

1. Read the schema first, before reading the document. Open schema.example.json in this kit and write down its two lists side by side: the three entity types it names (Service, Incident, Cause) and the two relationships, each with its required direction (a caused-by edge always runs from an Incident to a Cause; an affected edge always runs from an Incident to a Service). Note the required properties listed for each type. Treat this list as frozen for the whole run — nothing gets added to it mid-extraction.
2. Read the source document. Default to sample-input.md in this kit unless the user points you at a different file.
3. Draft candidate facts. For every statement in the document that looks like an entity or a relationship, write down a candidate in the form {entity_type, entity_properties} or {relationship_type, from, to}. Be generous at this stage — over-collect rather than under-collect.
4. Filter every candidate against the schema from step 1. Keep a candidate only if:
   • for an entity: its entity_type is one of Service, Incident, Cause, and its properties are a subset of that type's allowed properties;
   • for a relationship: its relationship_type is one of caused-by or affected, and the types of its from and to entities match what the schema declares for that relationship. Drop everything else. Do not rename a candidate's type to make it fit (e.g. do not relabel a Person as a Service) — a mismatched candidate is dropped, not reshaped.
5. Record every dropped candidate with the reason it failed (unknown entity type, unknown relationship type, or a from/to type mismatch). The on-call engineer's name and the Slack channel mentioned in sample-input.md are expected to be dropped here — that's the schema working as intended, not a bug to fix.
6. Emit the surviving facts as JSON shaped like schema.example.json's entities and relationships arrays, with each relationship tagged with a source_document property naming the file you read in step 2. Write the result to output.json in the kit's root (or the path the user requested) and also print it.
7. Print the drop list from step 5 alongside the JSON so a human can see what was excluded and why, without having to re-run the extraction to find out.