curate

name	curate
description	Interactive knowledge base curation for winze. Ingest sources, audit neighborhoods, check predictions, view KB health. Use when working on winze as a crew member or in interactive sessions.
allowed-tools	Bash(go *), Bash(git *), Read, Write, Edit, Glob, Grep, Agent
version	1.0.0
author	winze
argument-hint	<command> [args] — status, ingest <source>, audit [entity], predict, resolve, sensor

/curate — Winze KB Curation Skill

Route on the first argument:

/curate status

Show current KB health dashboard:

go run ./cmd/lint . 2>&1
echo "---"
git log --oneline -5
echo "---"
cat .sensor-state.json 2>/dev/null || echo "No sensor state yet"

Report: entity count, role types, lint findings, recent commits, last sensor poll.

/curate ingest <source>

Guided ingest of a source into winze. The <source> argument is a URL, DOI, Wikipedia article name, or file path.

Steps:

1. Fetch/read the source
2. Identify extractable entities and claims
3. Check schema fitness (do existing predicates suffice?)
4. Write the .go file with Provenance, entities, claims
5. Run quality gates: go build ./..., go vet ./..., go run ./cmd/lint .
6. Report findings and ask user to review before committing

Follow mirror-source-commitments: only encode what the source commits to.

/curate audit [entity]

Audit a specific entity's neighborhood, or the whole KB if no entity given.

For a specific entity:

1. Find all claims where entity is Subject or Object
2. Check Brief quality (is it substantive?)
3. Check cross-file bridges (does this entity connect to other ingests?)
4. Identify thin spots (claims that could benefit from additional sources)

For whole KB:

go run ./cmd/lint .
go run ./cmd/lint . --llm --llm-max-calls=10

/curate predict

Record a prediction about the KB or external signal. Prompts for:

• What is being predicted (natural language)
• Credence (0-1)
• Resolution date
• What signal would resolve it

Currently records as a typed claim pending the Prediction type addition to predicates.go.

/curate resolve

Check pending predictions against current state. Review any predictions whose resolution date has passed or whose signal has arrived.

/curate sensor

Run the Semantic Scholar sensor:

go run ./cmd/sensor --dir . --json

Show new papers found. For each paper matching a winze entity or concept, suggest whether it's worth ingesting.