Loading…
Use when extracting named entities and normalizing each to the org's canonical id + type code (a registry the base can't know) rather than raw surface mentions — so the same real-world entity resolves to one id downstream. Template: replace with your entity registry and codes.
| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| other-org-2 | ·→· | ⚠ Error | 237% | 0% |
| person-org | ·→· | ⚠ Error | 238% | 0% |
| person-initech | ·→· | ⚠ Error | 235% | 0% |
| two-prods | ·→· | ⚠ Error | 236% | 0% |
| person-london | ·→· | ⚠ Error | 235% | 0% |
“The skill is a benign template for entity normalization, providing clear rules for mapping named entities to canonical IDs.”
> Template skill. The rules below are a worked example (a fictional "Acme"). Fork this skill and replace them with your organization's real rules — the agent then applies your policy instead of guessing. The measured lift demonstrates the shape works.
Extract every named entity from the text into a JSON ARRAY of {type, id} objects. Use OUR type codes and canonical ids (the model cannot know these otherwise):
Initech -> "ORG-INITECH"; Umbrella -> "ORG-UMBRELLA"; any OTHER organization -> "ORG-OTHER".
Include only ORG/PER/PROD/LOC entities; deduplicate.