Which help center article to write next, from your Freshchat chats

Build me an app for the person who owns our help center. I am the one who decides what gets documented next, and right now I guess. I want the app to open on a ranked backlog of the questions our customers actually ask in Freshchat, one row per recurring theme, so my writing queue comes from real demand instead of my intuition.

Each row is a theme phrased the way customers phrase it, and shows how many distinct people asked, whether it is trending up or down against the previous equivalent period, three real customer verbatims lifted from the chats, and links through to the underlying Freshchat conversations so I can read the full exchange. Sort by volume by default, and let me filter to rising themes only. Clicking a row opens a detail panel with the full evidence list and every conversation behind the theme.

Mine the themes from Freshchat, and be careful here, because Freshchat has no account wide conversation list. The only route is to page List Users to enumerate contacts, call List User Conversations for each user, use Get Conversation to read status and timestamps so you can filter to the window I asked for, and then call List Conversation Messages on each conversation in range to read the actual text. Message lists default to 20 per page and cap at 50, so page them properly. Because this is expensive, never re-read every chat when the app opens. Give me a mining control where I pick a date range and start a run, show progress while it works, and persist the resulting themes, verbatims and conversation links so the backlog loads instantly from then on. A later run merges new evidence into an existing theme rather than creating a duplicate row.

Themes must be built from customer messages only. Filter out anything written by an agent or a bot before clustering, so agent replies, canned responses and greetings never create or inflate a theme. Rank by the number of distinct people who asked rather than raw message count, so one long back and forth does not outrank a question fifteen different customers each asked once.

Beside every theme show whether our Notion help center already covers it. Run Search by Title with the theme phrasing and its keywords, and Query a Data Source against our help center database for a filtered pass over published pages, then use Retrieve Page as Markdown on the closest matches to judge whether a page genuinely answers the question or only brushes past it. Label every row with one of exactly three states: covered, thin, or no article at all. When there is a closest match, show it on the row with a link, so I can decide whether to improve the existing page instead of writing a new one. Let me choose which Notion database is the help center in a settings screen.

Give me three actions on each row. Dismiss drops the theme from the backlog with an optional reason. Mark covered by hand lets me record that we already have an article the coverage check missed, optionally pasting the Notion link. Draft article kicks off the background agent.

Draft article starts a background agent scoped to that one theme. It reads the underlying transcripts through List Conversation Messages for the conversations behind the theme, works out what customers are genuinely stuck on and in what words, and writes a complete help center article in our house structure: the question stated the way customers ask it, a short direct answer up front, step by step instructions, and a closing section on what to do if it still does not work. It then files the article into our Notion knowledge base with Create a Page under the help center database, and writes the body with Update Page Content as Markdown. When it finishes it links the new page back onto the theme row along with its draft status, so I can see what has been drafted without leaving the app. Let me keep working while it runs, and let several themes draft at once.

Add a second view that flips the whole thing around: our published Notion pages listed against demand. Load the help center pages with Query a Data Source, match each one to the mined themes, and show how many customer questions map to it across the window. Sort ascending to surface published articles nobody ever needed, and descending to surface the articles people ask about constantly, which are the ones worth expanding or linking more prominently. Flag pages with no matching demand at all as candidates for retirement.

Persist everything inside the app: themes, evidence verbatims, conversation links, coverage states, dismissals, manual covered marks, and the Notion pages the agent created. A dismissed theme must never resurface in a later mining run even if customers keep asking it, unless I explicitly bring it back from a dismissed list. Keep the verbatims attached to every row permanently, because I need the customer's own wording when I sit down to write. Show the date range each theme's numbers came from, so the trend arrow is never ambiguous.