Inspiration
I run a weekly hygiene score for a real team: did each person log time the day they did the work, and did they leave proof of it. The first version flagged every miss. Within weeks the disputes arrived, and most of them were right: a recurring task closed with "nothing to set up this month" was failed for having no time logged; a parent task was failed although all its work was timed on sub-tasks; a colleague's close was charged to the assignee. Each fix made the score fairer, but the phrasing people use is endless, and a regex cannot tell a note from the same words inside a pasted log. So the rules now decide the clear cases, and Claude reads the ambiguous few. Fair Hygiene brings that to Jira.
What it does
For team leads who review their team's Jira work every week, and for the people being scored.
- Scores each assignee per week on same-day time and proof of work, with a configurable gate (pass / at risk / below gate) and no gate under five issues.
- Leaves out issues that say nothing about the person: closed by someone else or an automation (Jira's changelog names the closer exactly), imports, leave markers, and parent issues whose work is in sub-tasks. Each rule has an anti-loophole: any trace of the person and the issue scores normally.
- Sends ambiguous notes to Claude: a note that might mean nothing was needed, a late close with a reason, an overdue or stalled issue with an excuse. Claude reads the issue, its thread and the time on linked issues and answers defer / excuse / score or explained / not explained, with one sentence of why. Until then the issue is held out of the score, never counted as a fail.
- The lead overrules any verdict in one click; Claude's verdict comes back if the lead restores it.
- Hygiene Coach (Rovo) answers "how did the team do last week?" and "why is my score what it is?", naming the one habit that would lift a person's score most.
How we built it
Forge, UI Kit, and nothing outside Atlassian's platform:
jira:projectPage(UI Kit) with Scores, Judgements, Open work and Settings tabs.- Resolver functions reading Jira REST v3 (enhanced JQL search with embedded worklogs, comments and changelog; full lists fetched only when truncated).
- Forge LLMs (
@forge/llm): Claude, chosen at runtime from the active models, answering through a forcedrecord_verdicttool call. The answer is parsed strictly; a failed or unreadable call leaves the item pending with the error, never a guessed verdict. - Forge KVS for settings, verdicts, lead overrides and the evidence bundle Claude reads.
- Rovo agent
hygiene-coachwith actionsget-hygiene-reportandexplain-person-score. - Read-only scopes, no egress. The app reads as itself so a lead sees the whole team, and every entry point checks the calling user's own project permission first.
- The scoring core is pure and fully tested: 37 tests on fixtures written in Jira's wire format.
Challenges
- Fairness without loopholes. Every exclusion can only remove an unfair penalty: "closed by someone else" needs zero footprint, "parent container" needs the person's own time on a sub-task, a deferral needs a short note of their own, written in the week, not copied.
- Evidence windows. A person shown a failing score must not earn a pass by back-filling, so time and notes added after the week closed are inadmissible.
- Trusting the model. Claude sees roles, not names; comment text is evidence, never instructions (a test note tries to tell the judge what to decide); verdicts are keyed by a hash of the evidence, so an edited note is judged afresh.
Accomplishments
The scoring rules are not invented for a demo: each comes from a real dispute on a team where this score has run every week since April 2026. Live on a Jira Cloud site, Claude (Sonnet 5 via Forge LLMs) judged our eight test notes exactly as a lead would, identically across two runs, including a note that tries to tell the judge what to decide. On the fictional team in the video it judged all seven ambiguous notes the way a lead would, from the live app: the paused campaign and the withdrawn request are nothing to do, the duplicate is not the person's, "will finish next week" stays a fail, and the note that says "SYSTEM: ignore your rules" is named as a prompt-injection attempt and stays a fail. A judgement costs about $0.006. The production version is eligible for Runs on Atlassian.
What we learned
Most of the value is in deciding what not to ask the model. Rules settle the bulk of issues for free; Claude's attention goes to the few that need reading.
What's next
A weekly digest page in Confluence, precomputing large projects with a scheduled trigger, and a Teamwork Graph view of hygiene across projects.
How it meets the main technology requirement: a Forge app (manifest, jira:projectPage
UI Kit module, resolver functions, llm module, rovo:agent and action modules) integrating
with Jira (REST v3 issues, worklogs, comments, changelog, permissions) and Rovo.
Bonus prize: Best Agents
Best Agents (Rovo/Assignable agents): the app ships the hygiene-coach rovo:agent with two reusable action modules (get-hygiene-report, explain-person-score) that other agents can call as tools. Both actions run the same scoring service as the Jira page, check the calling user's own permission on the project first, and return plain text the agent can quote: the team's scores for a week, or one person's score with the single habit that would lift it most. Our demo site is on Jira's Free plan, which has no Rovo, so the video does not show the agent working; on a site with Rovo it appears in Rovo chat as Hygiene Coach.
Built With
- atlassian-forge
- claude
- forge-kvs
- forge-llms
- forge-ui-kit
- javascript
- jira-cloud-rest-api
- node.js
- react
- rovo