Inspiration
Teams already lose 25% of their time just searching for answers (Atlassian State of Teams 2025, 12,000 knowledge workers). The worse problem is finding an answer that is wrong.
Documentation drifts the moment work ships. In GitBook's State of Docs 2026, keeping docs in sync with the product is the #1 workflow challenge for 30% of teams, and for 43% of engineers. And 83% of busy documentation sites had at least one page flagged by their AI assistant as possibly outdated or incorrect.
Nothing turns red when a page goes stale. The runbook still says "page through PagerDuty" weeks after the team moved to Opsgenie. A new engineer follows it at 3 a.m. And now Rovo reads that page and repeats it, with full confidence.
Today's tools flag pages by age. Age is the wrong signal: "Founded in 2010" is old and true; "press the red button" was edited last week and is wrong. The real signal is the work your team just finished. That is what Driftcheck reads.
What it does
Driftcheck keeps your Confluence as true as your Jira.
The moment a ticket moves to Done, Driftcheck finds every page that work just made wrong, and tells the person who owns it.
- Proof, not guesses. Each warning shows the exact wrong sentence, the line from the ticket that proves it, and the corrected sentence.
- One click to fix. Apply fix rewrites the sentence and records "updated after OPS-2" in the page history. Still correct silences it for good.
- The right person, right away. Driftcheck comments on the finished ticket and @mentions each page owner.
- Catches what search can't. A page that says "open the PD app" shares no words with "PagerDuty → Opsgenie". Driftcheck still catches it, through the Teamwork Graph link. French page, English ticket? Caught, and fixed in French.
- Doesn't cry wolf. Rolled-back changes, history, old-but-true pages: left alone.
- Every space at a glance. A dashboard shows which pages are out of date and which are checked.
- Just ask Rovo. "Is this page still up to date?" "Which pages did OPS-3 make wrong?" The Driftcheck agent answers with proof.
How we built it
Forge only, Runs on Atlassian (no external calls):
- UI Kit:
confluence:contentBylineItem,confluence:spacePage - Trigger:
avi:jira:updated:issue, deduplicated per transition, daily cap in Forge storage - Rovo:
rovo:agent+ 2 reusableactionmodules - Forge LLMs: Atlassian-hosted Claude; long pages chunked and checked in parallel
- Teamwork Graph API: persisted Cypher queries on
content_entity_links_entity, with fallback to Jira remote links - Candidate selection: linked pages + keyword matches, then an LLM triage pass that picks tickets/pages by meaning (cross-language, abbreviations, synonyms)
- Precision guards: the model must quote the sentence verbatim and the exact words in it that are false today; both are string-checked against the page, otherwise the finding is dropped
Challenges we ran into
- False alarms cost more trust than misses. History ("in 2025 we moved to X"), generic wording ("the incident channel"), another instance (Paris vs. Lyon office), rolled-back changes and "incomplete but true" sentences were all flagged early on. Fixed with explicit rules plus the "quote the false words" guard.
- Words don't match meaning. French page vs. English ticket, "DS" = DocuSign. Solved with LLM triage instead of keyword-only matching.
- Big sites. On a site with hundreds of pages, keyword search drowned in noise ("engineers", "October"). Now: wide candidate pool, then LLM selection.
- Teamwork Graph access. Raw Cypher is restricted for apps; we moved to persisted queries owned by the app.
Accomplishments that we're proud of
Measured, not claimed: 49 adversarial test pages (changes made twice, abbreviations, other languages, rolled-back decisions, history next to current facts). Final version: 20/20 out-of-date sentences caught, 20/20 correct fixes, 1/21 false alarms. The fixes were tested on a second set written after the fixes, so we weren't grading our own homework. Every failure and fix is documented in the README.
What we learned
Drift detection is a precision problem. Requiring evidence on both sides (the ticket quote and the exact false words) did more for quality than any prompt wording. Keeping the human one click from the fix makes it usable.
What's next for Driftcheck
Bitbucket pull requests as a second source of truth, owner notifications inside Confluence, and cross-language detection in the automatic check.
Built With
- atlassian-forge
- claude
- confluence
- cypher
- forge
- forge-llms
- graphql
- javascript
- jira
- node.js
- react
- rovo
- teamwork-graph
- ui-kit