Sage

Agents ask. Product owners and their reviewers own the documents the answers come from. Design only the first half and you get a fast way to repeat whatever the documents say, out-of-date parts included.

The two halves meet in one loop, and the loop is the product.

Type the customer's question roughly as they asked it, get a few lines back. The sources sit under every answer, so checking is a click, not a search.

The sourcing is the part I would not trade away. An answer an agent cannot check is worse than the folder it replaced.

Accuracy is the adoption problem. One confident wrong answer in front of a customer and the agent goes back to the portal for good. So the thumbs-down asks three things: what kind of problem (incorrect or outdated, missing, unclear), which source, and optionally the sentence itself.

That becomes an issue for whoever owns the document. They land on the page with the sentence marked, the answer that was given and the agent's note, then mark it fixed or no fix needed. The status goes back to whoever raised it.

The first library mirrored the org chart: business line, a row of category tabs, then a grid of cards. It works when you already know where a document lives, and it breaks as the library grows. Every new category is another tab, every new document another card to scan.

People look for a document in one of two ways. By where they work: someone in HR does not care about Finance. Or by what kind of document it is: someone checking every live campaign wants Campaigns, whichever team wrote them. The new library pivots one tree between the two, with a count on every branch and a sortable table beside it.

I chose the tree because it scales. The document list will keep growing. A tree takes a new category as one more branch, and a table takes a hundred documents the same way it takes ten. Tabs and cards do not.

The first review flow gave every checker Accept and Reject at every stage. A rejection sent the draft back, a fix restarted the round, and a document could loop between maker and checkers without anyone knowing what was still blocking it.

I took Reject away from checkers. A checker either approves, or leaves a comment and does not. Nothing moves until every assigned checker has approved, then the head of the area gives final approval. They are the only one who can reject, which sends the document back to Draft.

Editing a live document starts a new version. The original stays published until the new one is approved, so agents never lose it mid-fix.

A checker with an open comment cannot approve. The button locks and says why. The maker replies and marks it addressed, the checker resolves it, and only then does Approve come back.

It is deliberate friction. Without it, a comment is a suggestion someone can approve straight past, and the document publishes with the question still open.

Figma first, for the handful of screens where the layout needed judgement. Then a DESIGN.md, then Claude Code building against it.

Generated UI drifts: ten sessions, ten spacing scales. The DESIGN.md writes the system down once, so the model works inside it. A few of its rules, as written:

Success is agents choosing Sage over the portal on a live call. There are no adoption numbers to share yet, so this is how I would show the loop is working: the share of flagged issues closed within a week, whether the same document keeps collecting thumbs-downs after it is fixed, and weekly active agents over time.