Viral Scout

Finds what is already working, so a reel can borrow a format — never a topic. Ends by putting picks in backlog/ideas.json, which reel-script reads.

No Apify, no API keys, no cost

scripts/find_viral.py reads each platform's own embedded JSON, the technique that replaced a $229/mo Apify scraper in app4 at 100% recall:

• YouTube — the search page embeds ytInitialData with a view count per result. Plain fetch, no browser, no quota. Sorted by view count server-side.
• Instagram — i.instagram.com/api/v1/users/web_profile_info returns recent media with video_play_count, given the public X-IG-App-ID header. This is per account, not search, so you curate whose numbers matter — better signal than hashtag roulette.

BROLL=.claude/skills/viral-scout/scripts
python3 $BROLL/find_viral.py \
    --yt "AI agent facebook ads" "shopify AI automation" --lang en \
    --ig shopify klaviyo triplewhale \
    --min-views 500000 --out work/candidates.json
python3 $BROLL/find_viral.py --yt "ИИ для интернет-магазина" --lang ru --min-views 200000

Run English and Russian as separate passes — the --lang header changes what YouTube returns, and Russian thresholds should be lower (a 200k RU video is the same signal as a 500k EN one).

The discipline: format, not topic

A raw run returns a lot of "I Tried AI Dropshipping For 7 Days". That is not a reel to make; it is a format that works. Never bring the user a list to copy the subject of. For each candidate, answer one question:

What did this video do in its first 3 seconds that made people stay?

Then bring the ones whose answer transfers to a first-party technical take.

Reject on sight, whatever the view count:

• get-rich / dropshipping / "make money with AI" — wrong audience entirely
• tutorials and step-by-steps — a different content type, not an opinion reel
• anything whose hook is a thumbnail promise the video does not pay off
• videos over ~10 minutes; the format does not transfer to 90 seconds

Keep when the video is: an operator telling a specific failure, a contrarian take on a tool everyone likes, a number nobody else published, or a teardown of something that broke in public.

Step 1 — Choose the queries

Read .claude/skills/reel-script/voice.md first. The beat is AI × e-commerce: agents running ads, analytics copilots, AI in checkout and support, platform automation. Not general AI, not general startups — those pools are enormous and almost none of it lands with a DTC founder.

Six to eight queries per run, mixing the two sides of the beat. For Instagram, keep a standing account list in backlog/ig_accounts.txt rather than guessing.

Step 2 — Shortlist

Present a table, ~10 rows, best first. Views alone is not the ranking — a 600k-view operator confession beats a 4M-view dropshipping video every time.

| # | views | platform | who | what the hook does | transferable? |

Say plainly which ones you would skip and why. A run where you recommend two of ten is a good run.

⏸ Gate — the user picks

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

Step 3 — Save with the lesson, not the link

A bare URL is worthless in a week. Each backlog entry records the format observation that made it worth keeping:

{"id":"v-20260831-01","kind":"video","status":"new","added":"2026-08-31",
 "url":"https://youtube.com/watch?v=...","views":612000,"platform":"youtube",
 "title":"...","why":"opens on the invoice, not the story — the number is the hook",
 "steal":"lead with the receipt on screen before saying a word",
 "topic":["ads","agents"]}

Append to backlog/ideas.json. reel-script Step 1 reads steal when shaping the hook.

Principles

• Format, not topic. Copying the subject makes you the fourth person with that take.
• An unknown view count never clears the threshold — it is excluded, not assumed.
• Two good picks beat ten mediocre ones; the backlog is a queue, not a hoard.
• Russian and English are separate runs with separate thresholds.