Skip to content

How to consume profiles efficiently

When reading a profile programmatically (especially from an AI agent), fetch only what you need. The manifest is a menu, not a preload list.

A profile can contain hundreds of files. Fetching everything for a simple question wastes bandwidth and context. Follow this order, stopping as soon as you have enough information:

StageFetchWhen to useCap
0Resolve base URL to the manifestAlways1 redirect
1profile.jsonldAlways, first1
2soul + expertiseVoice, opinion, critiqueBoth
3Works collectionPapers, years, venues, counts1
4Paper summaries (paper_summary)Evidence for claims≤8/turn, ≤20/session
5Full textOnly if summary can’t answer AND user named the paper≤2/turn, announce first
6grantsFunding questions (deep only)On demand
noneembedding_indexNever by a text agentnone

A consumer that fetches everything for a Stage-1 question wastes roughly 300x the bandwidth of a manifest-only read. The measurement behind that figure is in the consumer skill’s read-order reference.

When the manifest has a citations entry, use it to find relevant papers before fetching summaries:

  1. Check if the question names specific papers
  2. If so, verify they exist in the works collection
  3. Use the citation graph to find related papers
  4. Fetch summaries only for identified papers