How we tested Brook's reading
Before building the app, we checked the one part everything else rests on: given a saved post, can Brook tell what it is and what it is called? Run on 2026-10-09.
What we gave it
99 real posts people save, collected on 2026-10-09: 71 from TikTok and 28 from Instagram. Two Instagram reels could not be read (likely private or deleted), so 97 real saved posts were scored.
For each post, the model got only the platform, the caption and one picture, taken from the post's public preview without logging in: TikTok's embed data and Instagram's link preview. No video, no sound. That is what Brook is planned to read when you share a post to it; the app does not exist yet, so the test could not use it.
What it had to answer
- The type: recipe, place, product, workout, idea, or none of those.
- The name: the dish, the place, the product, the workout or the idea.
- The city, for places.
How it was scored
Each post was first labelled with the right answer by a stronger model, working from six frames of the video, its transcript and its caption; uncertain labels were reviewed, and 17 borderline posts accept either of two types (country trivia, for example, is both a place and none of those).
A type counts as right when it matches the label. A name counts as right when it refers to the same thing as the label, and, for a place, the city matches too. A separate model that was not one of those tested judged the names.
The result
Brook plans to use claude-haiku-5-5 from Anthropic, one of seven models we tried.
- It got the type right on 96 of 97.
- It got the name right, and the city for places, on 82 of the 90 posts that were a recipe, a place, a product, a workout or an idea.
We set the bar before the run: under 85% on types or under 70% on names would have stopped the project. It passed both.
What it does not cover
- Only the type, the name and the city were measured. The other details Brook is planned to fill in (what to order, the tip, when to go, price, time, ingredients) have not been tested yet.
- 97 posts is a small sample, gathered by us, not a random draw of everything people save.
- The right answers were labelled by a model and reviewed, not written by hand for all 99 posts.
- It tested the reading only. Brook itself is not out yet.