Answer a question from what is on the page
“What reference code is written in this task?”
The answer is on the screen, in a field people did not notice. Skate reads it and says where it found it.
Recorded on Vikunja, open-source, unmodified. Sped up.
What Skate does
- Reads the page.
- Finds the field.
- Answers with the value it saw.
Checked on the page: The answer, read from the page.
Proof
| App | Case | Task | Time |
|---|---|---|---|
| Vikunja (tasks) | case 11 | Answer a task detail from the page | 3 s |
| Gitea (dev) | case 09 | Answer an issue detail from the page | 3 s |
Benchmark run, 29 Sep 2026, on a self-hosted copy of the open-source app, graded by a separate checker that reads saved data. Single runs on apps we tune on every week. On a new app, expect about 4 in 10 on the first run (21 of 52 on two apps we had never seen), and about 8 in 10 once we tune it for that app.
Limit
Counting across several pages is weak: one counting case gave a wrong answer (Documenso 17).