Describe a task. See every screen of it.
Write what someone is trying to do and how you’d know it worked, and Variantly works through it from each perspective you name, keeping every screen it saw.
1 · Product design
Say what the task is for.
Two sentences hold every screen to account. The task is what someone is trying to do; “it worked when” is what should be true at the end.
- Someone wants to
- Invite a teammate with the right access.
- It worked when
- sam@demo-target.test is listed, and the admin knows it was sent.
2 · UX research
Choose whose experience you’re checking.
A perspective is a line or two about someone: what they know, how they get around, which account they use. Each gets its own attempt. Start with one and add more as questions come up.
- New administrator
- First week with the product assumed
- Signs in as admin test account
- Keyboard-only user
- Never uses a pointer assumed
- Signs in as member test account
3 · Usability and accessibility
Read every screen.
Each attempt comes back as numbered steps, each with its screen, accessibility tree and what a screen reader would announce. Wrong turns and dead ends are marked where they happened. AI proposes each issue; rules settle what they can.
Screen
Gone after 3 seconds
- Accessibility tree
generic "Invitation sent to sam@demo-target.test." (no role, no live region; removed after 3000 ms)- Transcript
button "Send invitation"heading level 1 "Team"Nothing about the invitation was announced.- Rule output
- The confirmation isn’t announced. Accessibility · set by rule
4 · Design engineering
Decide, then hand it off.
A person decides each issue: fix it, not sure yet, or working as intended. The ones you fix become fix briefs for Claude Code, Codex, Cursor or whoever makes the change.
Writes the fix brief. After the change, a recheck looks for what you intended.
Where: /app/team, #invite-flash (removed after 3000 ms; no role="status") Change: render the message in <div role="status">; remove the 3000 ms timeout; keep a Dismiss button. Recheck: message present at step 6; transcript includes "Invitation sent to sam@demo-target.test."
5 · Product
See that it landed.
After the change, the same task runs again from the same perspective and records what it saw: seen, not seen, or inconclusive. A person decides the issue, not a commit. Recorded tasks replay on a schedule, so a release that undoes the fix shows.
- Intended
- The admin knows the invitation was sent.
- Observed
- The confirmation vanished after 3 seconds.
- Rechecked
- Seen: at step 6 the confirmation stays, and it’s announced.
Use it between studies, not instead of them.
Research with people shows why they struggle, a few times a year. Variantly checks every release, accessibility included.
| Evidence | It shows you | It can’t show you |
|---|---|---|
| Research with people | Why people struggle, in their words | What changed in this week’s release |
| Product analytics | How many reached each step, and where they left | What was on the screen when they left |
| End-to-end checks | That the task still runs | Whether a person could tell what happened |
| Variantly | Every screen of a task from a named perspective, what got in the way, and a recheck | How many people would hit it. “Not sure yet” sends that to research |
What a review can and can’t tell you.
- An attempt isn’t a person. It’s one stated point of view working through the task. It can find friction a person might hit. It can’t tell you how many people hit it, or how they felt.
- A person decides every issue. AI proposes; you choose Fix it, Not sure yet, or Working as intended.
- Task-review issues never set a report row. Conformance report rows are set only by a rule over full coverage, or by a person with the evidence open.
- It runs where you say. Public pages, or signed-in areas with a test account you control. Credentials are sealed and opened only when an attempt needs them.
- A recheck records what it saw. After a change, the same task runs again from the same perspective, looks for the issue’s condition again and records what it saw: seen again, not seen, or couldn’t tell.
- Gaps are listed. Every review ends with what it didn’t cover, so a missing result is never read as a pass.