Nothing here is a black box. Every score has a breakdown, every draft has a reason, and every field says what it is for rather than what it is called.
A brand is you. A site is something you sell. You can have several of each, and they are scored separately because they are asking different questions of the same threads.
Paste a URL and it reads the page — title, description, headings — and proposes the terms. It shows you every line it drew them from. Nothing is saved until you have corrected it, because a wrong profile mis-scores every thread afterwards and you would never learn why the feed went bad.
Search terms are one or two words and decide what gets fetched. Problem phrases are full sentences and decide what scores. Neither can be derived from the other, so it collects both and labels which does what.
This is a rule because ignoring it broke a live crawl. A pasted site produced seven topics like “AI coding assistant context loss”. They saved. The crawler split them into single-word searches for “context” and “loss”, and found nothing at all.
Nine communities on a schedule. Threads arrive ranked, drafted, and explained — why this thread, why this angle — so a bad call is one you can see and overrule.
A built-in writer drafts the replies when no model gateway is configured, and the app says which engine wrote each one. Adding a key makes the drafts better; it is not what makes the product work.
Before you edit, not after. The counts sit above the thread: how many replies this hour, how many into this community today, how much of your last twenty promoted your own thing.
It never blocks the copy button. You are an adult with a reputation to manage and the tool does not own your account — it owes you the count, not a veto.
A scorer gets certain threads catastrophically wrong, and the ones it gets most wrong are the ones that look strongest to it.
Hacker News runs Who is hiring? on the first of every month. Each thread carries hundreds of comments and every comment is a job advert or a résumé. A résumé names more technologies in one paragraph than any real question does, so a keyword scorer reads it as the best match on the page. Measured on a live crawl:
13
of 100 Hacker News threads were recruiting
3
tied for first place in the feed
72
out of 100, the score they tied at
0
of them worth replying to
Replying is also the wrong act — that thread wants applications, not advice. So they are refused at ingest, before the write, and the run reports the count rather than dropping them quietly. The test matches the title only: a comment inherits its parent’s title, so one rule covers the monthly thread and everything under it, while a thread that merely discusses hiring is kept.
Eight sources need no account and a built-in writer drafts the replies. Keys make it better. They are not what make it work.