uberprompt infer

Find semantic edges between fragments that no uses declaration covers.

Usage

uberprompt infer

Reads all prompt and fragment files and sends their texts to gpt-5-nano. The model flags pairs of fragments that are related but have no declared uses edge. With --apply, the edges are written to apps/demo/edges.json as kind: "semantic".

Flags

Flag	Description
--threshold <n>	Minimum confidence score (0-1) to accept an edge (default: 0.7)
--apply	Write discovered edges to edges.json immediately
--json	Output results as JSON

Examples

Preview inferred edges without writing:

$ uberprompt infer

Inferred semantic edges:
  escalation-criteria ↔ refund-policy        (0.85)
  brand-voice ↔ output-format                (0.72)

2 edges found above threshold (0.70)
Run with --apply to write to edges.json

Apply with a higher threshold:

uberprompt infer --threshold 0.85 --apply

How it works

1. Loads all fragment texts from apps/demo/fragments/ and inline fragments from apps/demo/prompts/
2. Skips fragments with empty text (runtime input slots like )
3. Sends fragment pairs to the model, which scores how related each pair is
4. Filters results by the confidence threshold
5. With --apply, merges new edges into edges.json (existing edges are preserved)