A 30-second clip of a flickering dropdown tells an agent almost nothing. Without steps, expected behavior, and environment, the model fills gaps with plausible but wrong fixes.
A solid manual baseline works for most teams. Rewatch the recording once and fill a fixed template:
- repro steps with timestamps from the clip
- expected vs actual behavior in one sentence each
- browser, viewport, test account state, and feature flags
- console errors and failing request URLs
Paste that Markdown with the video link into the agent prompt. That alone removes most back-and-forth about what broke and where.
Keep clips under a minute and trim to one bug per file. If the repro needs login or seeded data, note the setup commands so the agent can reproduce locally.
The remaining gap is assembly cost. Copying timestamps, selectors, and logs by hand is slow and easy to skip under pressure. Capture-to-Markdown tooling can export the same structure automatically after one recording instead of typing it.
What do you include in UI bug reports for agents today, and what part do you still skip because it takes too long to capture?
Top comments (0)