Inspiration

Teams publish numbers every day: a refund policy that promises 30 days, a capacity plan that says 3 x 7 = 21 tons, a release note that says latency fell 40%. Nobody checks them, and an AI assistant that summarizes the page repeats them with total confidence. We built Humanities Helping Hands (HHH) as a checker that never pretends: every claim gets a label that says how it was checked, and "not checked" never means false.

What it does

HHH Claim Check brings that checker into Confluence and Jira:

  • Rovo agent, HHH Claim Checker. Ask "check the claims on this page" or "check this issue". It reads the page or issue, finds each checkable claim, and answers with the counts first, then every wrong claim with its reason and fix, then the proven ones, then what it could not check and how to check it by hand.
  • A button on every page and issue. A Confluence byline item and a Jira issue panel run the same check in one click and show each claim with a colored label: Proven, Wrong, Not checked, Open. Wrong claims come first, because they are the ones to fix.
  • The same checks for any AI agent, through Atlassian's MCP server. The app's rovo:mcp module exposes three tools, check text, look up a checked fact, and check a Confluence page by id, so a team's own agents can ask HHH before they repeat a number.

Every answer says exactly what was checked and what was not. A proven claim carries its receipt; anything the checker cannot settle is honestly labeled "not checked" with a next step, never guessed.

How we built it

  • Forge (Node.js 22 runtime): a Rovo agent (rovo:agent), five actions, a rovo:mcp module, a confluence:contentBylineItem and a jira:issuePanel built with UI Kit.
  • The page body is read with the Confluence REST API v2 (storage format) and the issue with the Jira REST API v3 (Atlassian Document Format), both as the signed-in user, so the app sees only what that person can see.
  • Each sentence is sent on its own (headings and titles are left out, since they name things rather than claim them), in batches, to HHH's public checking service (the same one behind humanitieshelpinghands.com and its public MCP server), and the answers come back in page order.
  • The checking service already existed and stays a private service behind a public API; everything Atlassian-facing (the agent, the actions, the MCP module, the panels) was built new on Forge for Codegeist.
  • Tests: the text extraction (storage format and ADF), the sentence cutter, the batching, the label counts and the app's own loading run offline with node --test.

Challenges we ran into

  • Rovo action inputs can be produced by the model, so the page and issue are always taken from the app's own context, never from model text.
  • A long page can exceed one check: the app checks the first 60 sentences and says plainly how many were left, instead of showing a partial count as the whole.
  • An agent loves to sound sure. Ours never upgrades a label, and never fills in from its own knowledge when a check fails.

Accomplishments that we're proud of

  • An agent that is allowed to say "not checked", and says why.
  • One checker behind three doors: the Rovo agent, the page and issue panels, and Atlassian's MCP server.

What we learned

Teams do not need an AI that is right more often as much as one that tells them which of its sentences were actually checked.

What's next for HHH Claim Check

  • Highlighting each claim in place on the page.
  • A Jira automation that checks an issue when it moves to Done.
  • Publishing to the Atlassian Marketplace.

Built With

Share this project:

Updates

Submission history