Topic radar

The user says a topic. You come back with the two or three things worth making a video about today, each with a reason it is worth making today. Not a search-results dump — a shortlist you would defend.

The scoring is done for you, in code, so it is the same every morning. Your job is the part code cannot do: read the primary sources and decide whether the top-ranked thing is actually a story.

Antigravity / Gemini: same steps, run the commands directly.

Step 1 — Run the radar

node scripts/research/topic-radar.mjs --topic "<what the user said>" --top 10

It fans out across four keyless sources, merges duplicates, scores, drops anything an earlier run already handed out, and writes brain/radar/<date>-<slug>.json + .md. It prints RADAR=<path> — capture that.

Flags worth knowing:

Read the ## Sources block in the output. It says what each source returned. A line like reddit — 47 items — RSS fallback … no vote counts is telling you that source's ranking is weaker than usual, and that is worth knowing before you trust position #1.

Step 2 — Read the primary sources

The radar ranks attention. It cannot tell a launch from a rumour. Take the top 3 and open what they actually point at:

• WebFetch the primary source — the announcement, the paper, the changelog, the repo README. Not a blog summarising it, and not the aggregator thread.
• Pull out the concrete, quotable facts: numbers, dates, names, what changed.
• If the link is a discussion thread, the item carries discussionUrl separately — read both. What the thread argues about is usually the better hook.

If a top item turns out to be a rewrite of something with no primary source behind it, drop it and say so. Ranking high on engagement and being true are different properties.

The full method, including screenshot evidence and B-roll, is in research-and-capture.

Step 3 — Report

For each story you kept:

**<the story in one line>**
Why today: <what changed, with the number or date that proves it>
Source: <primary url>   ·   score 0.69 · 1d old · hn + reddit
Angle: <the specific take — not "an overview of X">
Genre: news

Pick Genre from templates/VIDEO_GENRES.template.json — it decides which frames the script uses and in what order. news fits most radar output; technical-deepdive fits a paper or a release with real internals; listicle fits "five things shipped this week".

Sort by your judgement, not by the score. The score got you the shortlist; you rank it.

Step 4 — Hand off

The user usually wants a video next. That whole chain is daily-topic-video — do not re-run the radar for it, pass it the RADAR= path you already have.

When there is nothing

Some mornings a topic has produced nothing new. The script says so and exits with an empty RADAR=. Say that plainly. Offer --days 90, or a neighbouring topic, or --no-dedup to show what was already used.

Do not widen the window silently and present week-old news as today's. Do not fill the gap from memory — you do not know what shipped this week. An honest "nothing today" is a usable answer; a confident wrong one costs a video.

Notes

• No API keys. Reddit, Hacker News, GitHub and Google News all answer without an account. GITHUB_TOKEN is read if present but only raises a rate limit.
• Reddit's JSON API 403s unauthenticated clients on most networks; the source falls back to its Atom feed, which has the posts but no vote counts. Reported, not hidden.
• The news wire is Vietnamese by default (NEWS_RSS_LOCALE=vi|VN|VN:vi). Set en|US|US:en for an English-market agent.
• The seen-ledger lives at brain/radar-seen.json and records only what was handed out, not everything found — so a story ranked 40th on Monday can still surface on Tuesday.
• To see one source's raw output when a result looks wrong: node scripts/research/hot-sources.mjs --topic "…" --sources hn --json

What not to do

• Do not write the video from the radar entry alone. The title is an aggregator's title; the facts are in the primary source.
• Do not present a GitHub repo's star count as adoption. created: filtering means these are weeks old — stars that new measure attention, not use.
• Do not pass --no-dedup to make a thin morning look fuller. That is how the same story gets posted twice with a different voice.