By Vellora Labs · Updated October 11, 2026
A generated result can appear quickly and still be unusable. Study the journey from first visit to a result the person can inspect, correct, and use. Keep interface completion and output quality as separate observations.
Define the decision before the session.
Choose one uncertainty: whether people understand what input is needed, whether they can recover from an unsuitable result, or whether they recognize the product’s limits. Recruit people with the relevant task experience; document their prior familiarity with this product and similar AI tools.
Use test accounts and permissioned, non-sensitive inputs. Record product version, model or mode where visible, starting state, device, and any preloaded examples. Reset the environment between participants. Do not interpret a prefilled sample as evidence that a newcomer can prepare their own input.
Observe the first useful result in stages.
| Stage | What to observe | Keep separate |
|---|---|---|
| Before interaction | Ask what they expect the product to do and what they must supply. | Expectations before and after an explanation. |
| Input preparation | Observe missing context, permissions, uncertainty and workarounds. | Input difficulty versus interface navigation. |
| Generation | Capture waiting, errors, retries and what the person thinks is happening. | System latency versus participant effort. |
| Review | Ask how they decide whether the result is usable. | A finished generation versus a verified result. |
| Recovery and next action | Observe correction, regeneration, saving or export. | Assisted completion versus independent completion. |
A fictional session task
Teaching example, not a completed study: a teammate needs to explain one approved product capability in a short video. Supply a fictional feature brief and licensed demo assets. Ask the participant to prepare something they would feel comfortable sharing with that teammate. Do not specify buttons or a required editing route.
For this task, define a usable result in advance: it describes only the supplied capability, shows readable labels, and offers an appropriate next step. Observe whether the participant notices unsupported additions. A visually polished output that invents a feature does not meet that condition.
Record time without hiding what happened.
Keep the full elapsed time, active work, and system waiting in separate fields when you can observe them reliably. Mark interruptions and assistance. If the product never returns an output, record the failure and its context. Do not silently remove it from a report about the onboarding experience.
Use the finding to choose a change.
- Input expectation mismatch: revise the input example or explain required material, then retest unaided entry.
- Navigation problem: revise the interface and repeat the same goal-based task.
- Output error overlooked: improve the review step and inspect whether people catch the same error.
- Unclear next action: test the transition to editing, saving, export or sharing.
These are hypotheses for a next iteration, not measured conversion improvements. Retest with relevant participants and record the exact change. Keep disagreements and unsuccessful attempts in the evidence record.
Prepare the next pilot.
Build the task with the usability task builder, document attempts in the evidence log, and prepare materials with the video brief builder. TapVid is a product from the team behind Vellora and can produce product explanation videos; this method is a general study plan, not a published TapVid usability result.