Skip to content
Agent Orchestration
Skill

/jira-review

Review the pull request a Jira issue already carries, QA it on the deploy preview, and record the verdict on the issue. Insert into the prompt of a Jira review or QA column rule.

BOOST
From plugin
studio
41224 skills
Install
$ npx -y skills add decocms/studio --skill jira-review --agent claude-code

How it fires

How this skill gets triggered: by you, by Claude, or both.

  • Fires itselfAuto-invocation. Claude auto-loads it when your prompt matches the work.Auto-invocation is when the right skill fires by itself at the right moment, driven by a FLOW.md router and a hook, instead of you invoking it by name. It is the difference between a skill being installed and a skill actually getting used.Read the full definition →
  • You can call itInvoke it directly when you want it.
  • Slash command/jira-review

Context preview

The summary Claude sees to decide when to auto-load this skill.

Review the pull request a Jira issue already carries, QA it on the deploy preview, and record the verdict on the issue. Insert into the prompt of a Jira review or QA column rule.

SKILL.md

jira-review.SKILL.md
name: jira-review
description: Review the pull request a Jira issue already carries, QA it on the deploy preview, and record the verdict on the issue. Insert into the prompt of a Jira review or QA column rule.
disable-model-invocation: true

jira-review — review what the other run built

Insert this into the prompt of the Jira column your team reviews in. It is the second of two runs: an earlier column implemented the issue and opened a pull request (`jira-execute`), and this one judges it.

It is ordinary text once inserted — read it, cut what does not apply to your team, add what does.

---

You are reviewing someone else's change to a Jira issue. The issue itself is above, web links included.

Find what you are reviewing

  • The pull request is normally on the issue as a web link, put there by the run

that implemented it. If it is not, its branch is normally named for the issue key: `gh pr list --search "<the issue key>" --state all`.

  • If there is genuinely no pull request, do not write one. Report that on the

issue and stop.

Review the code

  • Check the branch out and read the diff (`gh pr diff <n>`, then the files

around it). Does it do what the issue asks; is it reachable from the surface the issue names; does it break something next to it.

  • **Do NOT implement the fix yourself and do NOT open a pull request.** A

reviewer that rewrites the change has reviewed nothing, and a second pull request on one issue is what a human then has to clean up. A small, unambiguous correction is a comment on the issue, not a commit.

QA it on the deploy preview

  • The deploy bot comments the preview URL on the pull request

(`gh pr view <n> --json comments`); the issue's web links may carry it too. If no preview exists yet, poll until it does. A first request may wake a hibernating build; retry until you get real HTML.

  • Exercise the behaviour and MEASURE it. Read the value back from the page

rather than inferring it from a name or a spec.

  • Compare against production, not only against the spec: the live site is the

"before" and the preview is the "after". A difference you cannot see on both is not evidence. If production is unreachable from this pod, say so in your report rather than passing the card on an untested comparison.

  • Capture before/after evidence with `qa-screenshot <url> <path>.png [--mobile]`,

on desktop AND mobile, and `Read` each file. A screenshot you never opened is not verification, and mobile is not desktop resized — plenty of these issues only reproduce on one of the two.

Record the verdict on the ISSUE

There is no card for anyone to read. A final message in this run is not a report.

  • `JIRA_COMMENT_ADD` posts your review (markdown, tables included). Open with

the verdict in ONE line — approved, or changes requested — naming the pull request and the head commit you reviewed (`gh pr view <n> --json headRefOid`), then the evidence, then the findings, each naming a file and line or a screenshot. The commit is what tells the next reader whether anything landed after your review.

  • A finding is a defect you OBSERVED, not a preference. Say what you did, what

you expected, and what happened instead. If you found nothing, say that plainly rather than padding the list: an invented finding costs more than a missed one here.

  • To show evidence, write each screenshot to `/app/org/output/<name>.png` and

reference it in that comment as `![what it shows](/app/org/output/<name>.png)` — it is uploaded to the issue and rendered inline. A before/after table of two images reads best.

  • `JIRA_REMOTE_LINK_ADD` adds the preview you reviewed (`key: "preview"`) if the

implementing run did not already put it on the issue.

  • `JIRA_ISSUE_GET` re-reads the issue before you report — a person may have

edited the card, or answered a question on it, while you worked.

  • `JIRA_ISSUE_TRANSITION` carries the verdict: on a pass, the column your team

hands reviewed work to; on a fail, BACK to the column the implementing run works in. Do this LAST, after the comment is on the card — a transition with no comment beside it is the one nobody can act on. Then **re-read the issue** and confirm it landed where you meant — a transition's advertised destination is not always where it puts the card, and the failure is silent.

  • If you cannot tell which column is which, say so in your comment and leave the

issue where it is. A card parked with an explanation is recoverable; one moved into the wrong column reads as a handoff nobody made.

Read more
Ships withstudio

Open-source control plane for your AI agents. Connect tools, hire agents, track every token and dollar

Get the whole plugin

Other skills on studio.