uberprompt propose

Turn unprocessed lessons into pending proposals.

Usage

uberprompt propose [--dry-run] [--model <m>]

Reads lessons that have no processedAt timestamp and generates prompt-change proposals. Each lesson is matched to target prompts by the targeting ladder, in order:

1. Lineage: the prompts whose traces produced the lesson (lesson.appliesTo)
2. Catalog reasoning: an LLM reads the full prompt catalog and picks other prompts the lesson applies to
3. RAG: vector search of the lesson embedding against prompt description embeddings

Proposals are written to the proposals collection with status "pending". Identical pending proposals are skipped. Proposals for the same prompt are grouped, so one approval produces one version bump.

Flags

Flag	Description
--dry-run	Print proposals without writing them to the database
--model <m>	LLM to use for catalog reasoning and rewrite generation (default: gpt-5.1)

Examples

Generate proposals from all unprocessed lessons:

$ uberprompt propose

Processing 2 unprocessed lessons...

Lesson: "never promise a credit amount before checking the delivery record"
  lineage: refund-checker
  catalog: escalation-writer (LLM match)
  Filed 2 proposals

Lesson: "always include order number in escalation context"
  lineage: triage-router
  Filed 1 proposal

3 proposals filed

Preview without writing:

uberprompt propose --dry-run

Prerequisites

Requires lessons in the database (produced by uberprompt learn). Set MONGODB_URI in your .env file.