comment-bait-reel
Build a "Comment X for the links" lead-magnet reel (9:16, ~14s) from the creator's OWN clips + REAL YouTube tutorial thumbnails. Proven format: [b-roll + "POV: you want X" caption] -> [black + "Watch
SKILL.mdskills/content-and-reels/comment-bait-reel/SKILL.md
GitHubcomment-bait-reel
A lead-magnet reel that converts a scroll into a comment, then a comment into a DM. The on-screen video is only half the system — the other half is the keyword funnel that delivers the promised links.
The format (decoded from the viral reference)
[B-ROLL + POV caption] → [BLACK + "Watch this" + real YouTube card] ×3
↓
[CTA: Comment "KEYWORD" / it's free]
- Each beat is short: b-roll ~1.6–1.8s, card ~2.0s, CTA ~3.3s. Total ~14s.
- B-roll = the creator (face/desk), graded, with a top dark bar + white "POV: you want X" caption. First b-roll is the HOOK — highest-stakes line, most cinematic frame. The hook frame decides reach; spend the most effort here.
- Card = pure black, big "Watch this", a real YouTube thumbnail framed white, real title + real channel name. NO invented stats (see hard rules).
- CTA = black, "Comment "KEYWORD"" + a punchy 2-word promise ("It's free") + one subline + a pulsing "comment below" nudge near the bottom.
- Cuts are hard, each segment opens with a 0.07s white flash (flash-cut) and cards get a slow zoom-in. That reads as "transitions" without fragile xfades.
Inputs you need before building
- Creator clips — 1–2 clean, caption-free vertical windows of the creator
(a tight intense close-up = best hook; a talking-head or desk shot = body).
Probe with
cropdetect(see gotchas) to find caption-free seconds. Letter- boxed clips with burned captions: crop the clean video band and blur-fill. - 3 real YouTube tutorials on the topic — get real
video_ids (research agent). Thumbnail =https://i.ytimg.com/vi/<id>/maxresdefault.jpg(fallbackhqdefault.jpg). VERIFY each id resolves and the thumb downloads (>3KB). Reject dead/private/tiny. - Keyword + lead magnet — the exact comment word (keep it the topic, e.g. CLAUDE) and what gets DM'd. Wire the DM via ManyChat keyword or GHL.
Build
The engine is assets/build.template.sh. Copy it into a working dir next to the
source clips, edit the 3 b-roll blocks (clip path + -ss/-t + POV text), the
3 make_card lines (thumb + title + channel), and the CTA keyword. Then:
bash build.sh # writes reel.mp4, regenerates everything idempotently
It builds 7 segments → concats with the concat demuxer (stream-copy, instant). Output is silent by design — add trending platform audio at upload (boosts reach + you avoid shipping unlicensed music).
HARD RULES (learned the hard way)
- No fake stats. Do NOT put invented view counts / "X weeks ago" / durations on the cards. The videos are real but the numbers aren't — fabricating stats about real third-party channels is a no-fake-claims violation AND collapses credibility the moment a viewer clicks through. Card = real title + real channel name only. (This was the #1 critic ship-blocker.)
- Hook is the whole ballgame. A bright low-effort selfie as frame 1 = scrolled. Lead with the most cinematic/intense creator frame + a high-stakes caption ("POV: you're scared AI will replace you" > "POV: you want to learn X").
- Keyword = the topic. The comment word should match what's delivered (CLAUDE for Claude links). One unambiguous word.
- Real thumbnails > designed cards. Real recs look like genuine tips. Verify each thumbnail visually before building (Read the jpg).
- Creator continuity. Carry at least the first two b-roll beats with the actual creator, not a stranger/stock — continuity sells the "I'll send you MY list" promise.
ffmpeg gotchas on Windows (gyan build)
- Font path colon breaks the filtergraph.
C\:/Windows/Fonts/...does NOT escape on this build, and fontconfig has no default config. FIX: copy the TTFs to a relative, colon-free dir (fonts/arialbd.ttf) and reference that.assets/fonts/ships arial / arialbd / ariblk. - Colons inside
text=break parsing even when single-quoted. "POV:" / clock values die. FIX: escape the colon as\:inside the text value. zoompandis frames-PER-input-frame. Feeding a looped image with-t 2(=60 input frames) ×d=60= 3600 frames = a 2-MINUTE clip. FIX: feed ONE frame (-loop 1 ... -frames:v 60) and letd=60generate the 2s.- concat demuxer paths are relative to the list file. If
list.txtlives insegments/, entries must bes1.mp4, NOTsegments/s1.mp4. - Always pass
-y(engine bakes it into$ENC+ does a cleanrmfirst) or ffmpeg blocks on the overwrite prompt in non-interactive shells. - Find the clean video band in a caption-burned letterboxed clip:
ffmpeg -i in.mp4 -vf cropdetect=20:2:0 -f null -(run at default loglevel —-v errorHIDES cropdetect output). Note: burned caption text counts as content, so the reported top edge includes the caption; the clean video is the lower sub-band — eyeball a frame to split them, then blur-fill the band.
Ship checklist
- Build → extract a frame per beat → Read them (visual QA).
- Independent critic pass (no self-grading): hook strength, card realism, caption legibility, CTA pull, topic coherence, no-fake-claims. Threshold 80.
- Wire the keyword funnel (ManyChat/GHL) BEFORE posting — the DM must fire.
- Post silent; add trending audio in-app. Pin a first comment with the keyword.