Read the skill at https://ai-skills-vault.pages.dev/skills/ai-video-editing-reels.md and install it as a skill in your agent. Follow it exactly, including its checklists. Then ask me the questions it requires and do the work.
The difference between a 2,000-view video and a 200,000-view video is usually the first 2 seconds and the cut rhythm. This skill engineers both.
When to use this skill
Producing Reels/Shorts/TikTok from raw footage, a screen recording or an AI avatar.
Turning long-form content into a series of shorts.
Diagnosing why a video underperforms.
Stage 1 - The retention architecture
Every short is built in five blocks. Get the seconds right:
Block
Time
Job
Hook
0.0-1.5s
Stop the scroll - motion, claim, or a visual oddity
Setup
1.5-4s
Promise what the viewer gets, create an open loop
Payload
4-20s
Deliver in 3-5 beats, one idea each
Twist / proof
20-30s
The thing they did not expect, or the evidence
CTA + loop
last 2-3s
Ask once, then cut so it loops to the hook
Under 30 seconds is the default. Up to 45 s only if the payload genuinely needs it.
Stage 2 - Hook formulas that work
Contradiction - "Your phone is why your 3D looks fake."
Number + outcome - "3 settings that doubled my render quality."
Direct callout - "If you edit on your phone, stop doing this."
Mid-action open - start inside the motion, no intro.
Result first - show the finished thing, then explain how.
Mistake reveal - "I wasted 6 hours so you don't have to."
Never open with a logo, a "hey guys", or an introduction. Those are three separate scroll triggers.
Stage 3 - Scripting (write it, do not improvise)
Format: [SHOT] | on-screen text | voiceover line | duration.
Rules: one idea per line, under 12 words per line, spoken at ~150 words per minute, read it aloud once and cut every word you stumbled on. Put the payoff at 60% of the runtime, not at the end - retention curves collapse after the payoff, so end fast.
Cut rhythm: a visual change every 1.5-2.5 s in the payload. A cut can be a new angle, a zoom, a text pop, a b-roll insert, or a jump cut. Something must move.
J-cut audio: start the next line's audio 4-6 frames before its picture. It removes the stitched feel.
Punch-ins: scale 100% to 108% on emphasis words; never more than 3 per 10 s.
B-roll: 1 b-roll per 5 s maximum; every b-roll must illustrate the words being spoken at that moment.
Speed ramps: 1.0x to 1.15x on transitions only. Slow motion needs 60 fps source; 30 fps slowed looks like a slideshow.
No transition packs. Hard cuts and one whip pan or flash. Everything else reads as 2016.
Stage 5 - Captions
Burned-in captions are mandatory - most viewers watch muted.
3-6 words per line, 2 lines max, centred in the middle third of the frame (safe from UI overlays).
Font: bold, heavy, high contrast; white with a 2-3 px dark stroke, or a solid accent box behind key words.
Highlight 1-2 keywords per line in the accent colour. Never highlight everything.
Keep captions inside the safe area: 1080x1920 frame, keep text within x 90-990, y 260-1500.
Stage 6 - Sound design
Sound is 50% of perceived quality.
Music: 90-110 BPM for tutorial, 120-140 for hype. Duck the music to -18 dB under voice; voice sits at -6 to -3 dB peak.
Cut on the beat: place cuts on kicks for the payload section.
Foley: whoosh on transitions, click on text pops, sub-bass hit on the hook. Keep all under -20 dB.
Normalise the final mix to -14 LUFS (platform standard). Loudness mismatch is why some videos sound "cheap" next to others.
Never use trending audio with a claim you cannot legally back; keep a licence note.
Stage 7 - Export specs
Platform
Frame
fps
Bitrate
Notes
Instagram Reels
1080x1920
30
10-12 Mbps
under 90 s, cover frame set manually
YouTube Shorts
1080x1920
30-60
12 Mbps
under 60 s for the Shorts shelf
TikTok
1080x1920
30
10 Mbps
avoid watermarks from other apps
Safe zones: keep text 250 px from the top and 320 px from the bottom (platform UI).
Stage 8 - Repurposing matrix
One long video becomes many shorts:
Pick the 5 most quotable 20-second moments.
Each becomes one short with its own hook written fresh - never reuse the long-form sentence as a hook.
Re-cut the first 3 seconds differently per platform; the same hook fatigues fast.
Repost the best performer after 3 weeks with a new hook and cover.
Keep a swipe file of hooks that worked; the hook is the reusable asset, not the video.
Stage 9 - Diagnose underperformance
Symptom
Likely cause
Fix
High views, low completion
Payoff too late
Move the payoff to 60%
Low views, good completion
Weak hook / bad cover
Rewrite first 1.5 s, change cover frame
Views stall at 3 s
Setup promises nothing
Add an open loop in the first line
Comments say "too fast"
Cut rhythm under 1 s
Lengthen beats to 2-2.5 s
Looks amateur
Flat audio, no captions, wrong bitrate
Normalise to -14 LUFS, add captions, re-export
Quality checklist
Hook lands in under 1.5 s, no logo or intro
A visual change every 1.5-2.5 s in the payload
Captions burned in, 3-6 words per line, inside safe zones
Music ducked, mix at -14 LUFS
Payoff at 60% of runtime
Loops cleanly back to the hook
Exported at platform spec, no watermark
Cover frame chosen deliberately
Copy-ready agent prompt
Read the skill at the URL above. I will give you a topic and my raw material. Produce: (1) three hook options using different formulas, (2) a full shot-by-shot script table with durations, (3) the exact CapCut edit steps with cut timings, (4) caption text split per line with the keyword to highlight, (5) export settings for my platform. Then run the quality checklist on your own output and fix anything that fails.