Topic Scout

Finds things worth having an opinion about, and rejects the rest. Ends by putting picks in backlog/ideas.json, which reel-script reads.

The beat

AI × e-commerce, not general AI and not general startups. Agents running ads, analytics copilots getting numbers wrong, AI in checkout and support, platform automation, what breaks in production at an operator's scale.

Read .claude/skills/reel-script/voice.md first. The audience has a P&L. They do not care about model architecture; they care what it costs them when it is wrong.

✓ Meta ships agent tooling for ad accounts      ✗ new LLM benchmark
✓ an operator's agent broke something in prod   ✗ YC batch roundup
✓ a platform publishes peak-season numbers      ✗ GPU funding round
✓ an analytics copilot reports a wrong figure   ✗ model release notes

Step 1 — Sweep

WebSearch across the last 7–14 days. Six to ten queries mixing both sides of the beat, plus the platforms by name (Shopify, Meta, Klaviyo, Amazon Seller Central, Google Ads). Recency is a fact about the fact — anything older than a month needs a reason to still matter.

Also sweep practitioner posts on X and LinkedIn where an operator describes something that actually broke. First-hand beats reported: the strongest reel this pipeline has produced was built on a four-day-old tweet.

Step 2 — Judge reel-ability, harshly

Most news cannot become a reel. A candidate needs at least three of:

1. A specific number — not "many teams", an actual figure with a source
2. A named party — a platform, a company, a person on the record
3. A conflict — someone is wrong, or two credible parties disagree
4. A first-party hook — he has shipped, measured or broken something related
5. Recency — it happened, or became visible, in the last two weeks

Fewer than three and it becomes a summary read aloud, which is the failure mode this whole pipeline exists to avoid.

Also reject: vendor announcements with no independent number, opinion pieces with no data, anything whose only angle is "this is interesting".

Step 3 — Shortlist

Present ~8 rows, strongest first:

| # | what happened | source + date | why it's reel-able | the angle he'd take |

That last column is the work. Not the topic — the position, in one sentence someone could argue with. If you cannot write it, the item is not a candidate and does not go on the list.

Flag anything shaky yourself: a vendor's stat about its own category, a single outlet with no primary source, a number that cannot be verified.

⏸ Gate — the user picks

They read, they choose. Never add to the backlog without an explicit pick.

Step 4 — Save the angle, not the headline

{"id":"a-20260831-01","kind":"article","status":"new","added":"2026-08-31",
 "url":"https://...","publication":"Shopify","date":"2025-12-02",
 "title":"Merchants hit $14.6B over BFCM",
 "angle":"BFCM is the first peak season running on code nobody on the team wrote",
 "opponent":"operators wiring agents into ad management for peak season",
 "facts":[{"value":"$5.1M per minute at peak","url":"https://..."}],
 "topic":["bfcm","agents"]}

Append to backlog/ideas.json. reel-script Step 1 reads angle and opponent and skips straight to research.

Principles

• The angle is the deliverable. A link with no position is not an idea.
• Three of the five reel-ability tests, or it does not go on the list.
• Say which ones you would skip and why — a run recommending two of eight is good.
• Primary sources over reporting; reporting over blog summaries; never a summary of a summary.
• First-hand practitioner accounts beat trade press, because nobody else has them.