Generated from
smart-base/skills/content/authoring-who-smart-guidelines/recommendation-candidates.md— do not edit here.
recommendation-candidates
Step 1 of recommendation-extraction, for one
source at a time. That skill holds the judgement — what counts as a
recommendation, and implementation guidance that prints practices rather than
recommendations (its Part 1). This one is the mechanical step around it.
Input and output
- Input:
input.schema.json - Output:
output.schema.json
In: an ingested library/<id>/ entry and, on the second pass, the agent’s
l1-recommendation-curation/v1.
Out: the candidate work list; once curated, the L1 3.0 graph (l1.json,
every node inferred), the proposed exclusions and the warnings.
Steps
- Run the Tool with no curation: it writes the work list, every normative sentence with its hints, nothing decided.
- The agent writes
curation.yaml, one decision per location, following recommendation-extraction Part 1. - Run the Tool again with the curation: it applies the gates and writes the graph. A refusal (a statement that is not verbatim, an unknown location) goes back to step 2.
Tool
l1-recommendation-candidates (smart-base/tools/l1-recommendation-candidates/l1-recommendation-candidates.ts):
bun run smart-base/scripts/l1-recommendation-candidates.ts --entry <library entry> [--curation <curation.yaml>] --out <dir> [--check]