Filling a collection

You can build a new collection from entities already on your account, in other collections or in the Inbox, without researching them again. Ask your assistant to do it in one of two ways:

  • Name the members when you know which entities you want.
  • Describe what belongs when you know the rule but not the list.

Naming the members

Ask your assistant to make a collection and list the entities you want in it. The entities aren’t copied: they are the same entities, with the same claims and sources, now also shown in the new collection.

You can also say which columns you want and what each one means. Mark any column whose value identifies the thing rather than describes it, such as a website, a ticker or a DOI. That lets later writes recognize an entity even when it comes in under a different name.

Describing what belongs

Describe the rule instead of the list, for example: European battery storage companies with utility customers, books you rated four or higher, or suppliers in the EU with a published audit. Your assistant finds the closest matches among the entities on your account, reads each one, and adds the ones that fit.

Keep in mind:

  • Each entity is judged only on what you’ve recorded about it. No new research is done and nothing is fetched, so a rule about something you never recorded will find nothing.
  • It runs once. The result is a set of rows you can edit, not a saved filter. Entities you record later aren’t added automatically; ask again to include them.

Writing a clear description

The description is read literally, and a loose one is read broadly. A vague description can give you a very different collection from the one you meant, not just a few extra rows.

For example, here are two ways of asking for the same collection: 

If you ask for:  “organizations developing frontier AI models”
you get:         31 companies. Robotics and world-model companies are
                 included, because their records say they train large models.

If you ask for:  “labs training general-purpose language models. Not
                 robotics, world model, video or image model companies.
                 OpenAI counts; Figure AI doesn't.”
you get:         12 companies, all of them the kind you meant.

To get the collection you want:

  • say what counts
  • say what doesn’t
  • give an example of each

This only matters for rules that need a judgment. A rule your data can check directly, such as a rating, a funding total or a region, doesn’t need it.

Reading the report

Treat the result as a first draft and read it. The report lists:

  • what was added, and how confident the match was for each one
  • how many entities were considered and didn’t fit
  • entities it couldn’t place, because nothing you’ve recorded about them addresses the rule

The last group is the most useful. It shows where your data is missing the answer. To fix it, record the information the rule asks about, or add the entity yourself if you already know it belongs.

Everything here is easy to undo. Ask your assistant to take any row out of the collection (the entity and everything recorded on it stay where they were), or to hide any columns the run added. Fixing what went wrong covers the other repairs.