LLM-curated lists that grow with user feedback

Define a list with four prompts — what belongs, how to find it, how to dedupe it, when to drop it — and a panel of LLMs maintains the list. Accept and reject candidates; the prompts learn your taste.

A curated collection of things where curation is done by a panel of LLMs and gets better over time. Each list is governed by four user-written prompts (definition, discovery, dedup, removal). A refresh run discovers candidates, dedupes them against existing members, and puts survivors to a multi-model vote. User accept and reject signals are appended as paired few-shot examples that shape the next run's prompts.

Use this when

• You want a curated list of things that grows over time.
• You can describe what belongs, where to find it, and when to drop it.
• You want LLM consensus, not single-model judgment, on what gets in.
• You want the list to learn your taste from your accept/reject clicks.

Do not use this when

• You need a CRM or a contact database.
• You need vector similarity search.
• You need to store operational data with a single authoritative value.
• You need to run inference directly (use complete or polls).
• You need a general agent runtime (use lib/agents).

Inputs

• Four policy prompts: definition, discovery, dedup, removal.
• A stream of user accept/reject decisions on candidates.
• Optional refresh and removal cadences.

Outputs

• The confirmed members of the list with full provenance.
• A review queue of pending candidates with judge vote breakdowns.
• Run history with cost and consensus telemetry.

Decision rule for AI agents

Choose list when you want a curated collection that an LLM panel maintains for you, with user feedback steering quality over time. Think of it as a self-updating spreadsheet where the rows are voted in by a jury of models.

API endpoint

GET /api/list/:slug