pm-uat-review
Use when a task is in pm_uat - verify every acceptance criterion and the human's comments yourself in the browser, on a device, or over a loopback HTTP request, record a verdict per criterion, then mo
SKILL.mdcatalog/agents/product-manager/skills/pm-uat-review/SKILL.md
GitHubPM UAT Review
Overview
pm_uat is the PM's own gate, not a rubber stamp on QA's work. You have no code-reading tools here — codebase_search, grep_code, get_repo_tree, read_file and the rest are gone in this column — so your verdict can only come from evidence you produced yourself this run: a browser walk-through, a device walk-through, or a direct HTTP request. There is no backend exception: a criterion with no screen still gets your own executed evidence, never a pass on QA's word alone.
1. Inputs
- The task's ORIGINAL description and every acceptance criterion.
- The human's requirement comments on the task (injected as "The human's requirements written on this task"). Where they disagree with the original description or a criterion, the comment wins — and what it asks for needs evidence like any criterion.
- QA's evidence: the
review_criterionnote on each criterion, andlist_test_casesforpassedcases matchingcriterion_id. Both are a cross-check, never a substitute — you still walk every criterion yourself, covered or not.
2. Get a URL — never production
First one that applies: get_task_preview — a preview that is ready AND built_from_pr_head (its open_url); else get_deploy_target with env="stage", if this repository has a stage environment recorded; else start_task_preview and use its url once one appears (call it again if the URL isn't there yet — it will not restart a preview already starting).
3. Walk each criterion
UI-observable criteria: browser_navigate → browser_wait_for → browser_fill/browser_click through the flow, browser_screenshot as evidence. Check at phone (browser_set_viewport {device:"mobile", width:360}) and desktop at minimum for anything visual.
API-only criteria (no screen to walk, usually task_type: "technical" or a backend criterion on a task/bug): boot the task branch with start_task_preview, then use http_request against that preview's http://127.0.0.1:<port> URL (the tool refuses anything that is not this machine, so a stage or production URL is never an option for it; a read-only stage check is fetch_url) to exercise exactly what the criterion asks: happy path, invalid input, unauthorized/forbidden, whatever the criterion names. This is not optional and not QA's job to cover for you — every API criterion gets your own request in this run.
Evidence note format (cite it in that criterion's own review_criterion note):
- UI:
Observed: <exact text/value/state> at <url> [360px|desktop|device] — <what the screenshot showed> - API:
Observed: <METHOD> <path> [+ body if relevant] → <status> <content-type>; <response excerpt that proves or disproves the criterion>
A note that only says "verified" or "works" is a Red Flag — it has to name what you saw.
4. Mobile path
The PM holds the mobile_* tools — use them on a mobile repository, mirroring QA: mobile_launch_app → mobile_wait_for → mobile_tap/mobile_type_text/mobile_swipe → mobile_screenshot/mobile_read_ui → mobile_release_device, always last, whether the walk-through passed or failed. Check portrait and landscape (mobile_rotate) at minimum per UI criterion. If any mobile_* call reports the device in use, stop — the run is parked and resumes by itself; never retry or guess at a verdict without the device. mobile_launch_app installs the build registered on the repository's deploy target, not necessarily your exact task branch — say in the note which build you observed if that distinction matters to the criterion.
5. Stakeholder-eyes pass
One walk of the original request as the requester would make it, independent of the criteria list. Look for:
- wrong content language (not the product audience's language)
- placeholder text (
lorem,TODO,{{…}}) - a broken image or a bare "?" icon
- horizontal overflow (the
browser_set_viewportreport lists overflowing elements) - a missing empty or error state the request implies but no criterion named
6. Classify every finding
- Breaks a criterion, a human comment, or the original request → a gap → reject the criterion(s),
need_revision. - A new idea not asked for →
create_board_taskin backlog (priority low, same repository/project). Never a gap, never blocks approval.
7. A human comment supersedes a criterion
Use cancel_criterion with a reason quoting the comment, then verify the comment's own version of the behaviour and cite it in the note of the criterion closest to it.
8. Re-entry after need_revision
Re-walk the previously rejected criteria first, then all the rest — moving goalposts is not allowed in either direction. A new gap found on re-entry must trace to a criterion, a human comment, or the original request; otherwise it's a backlog task (step 6), not a reason to bounce the task again.
9. Output
- Pass: every criterion has your own executed evidence and is approved via
review_criterion. Move tohuman_uat. Write no comment — the move and the approved criteria are the verdict. - Fail: reject the failing criteria via
review_criterionwith expected vs actual, one numberedadd_task_commentgap list, move toneed_revision.
Never approve by reading code. Never approve off QA's notes, a passed test case, or the diff alone — "I looked at the code and it looks good" and "QA already tested this" are both forbidden; only evidence you executed yourself, this run, counts.
Worked Example
Task: "Add task export endpoint" (task_type: "technical" — skips pm_uat, routed to human_uat by QA directly) vs. "Add export button to the project board" (task_type: "task", UI-facing):
- AC: "Given the board, When I click Export CSV, Then
<project>-tasks.csvdownloads and the row count matches the board." →browser_navigateto the preview,browser_clickExport, confirm the download, screenshot. Approved, note: "Observed:acme-web-tasks.csvdownloaded at <url> desktop — 42 rows, matches board count." - AC: "Given the export is in progress, When I click Export again, Then the button is disabled and no second file downloads." → walked, button stayed enabled, a second file downloaded. Rejected, note: "Expected: disabled during export. Actual: button stayed clickable, two files downloaded." One gap-list comment,
need_revision.
Common Mistakes
- Treating a
passedQA test case as enough — you still have to run it yourself. - Skipping an API-only criterion because "nothing to click" — use
http_request. - Writing "verified" in a note instead of the exact observable.
- A "PM UAT passed" comment on a clean pass — the column wants no comment at all.
- Editing
acceptance_criteriahere instead ofcancel_criterionwhen a human comment supersedes one. - Re-litigating an already-approved criterion on re-entry instead of focusing on the rejected ones.
Red Flags
- A
review_criterionnote that says only "works" or "verified". - An approval with zero tool calls in this run's transcript.
- A gap list comment with no expected-vs-actual.
- A new idea folded into the gap list instead of a new backlog task.