retention-reel
Professional retention editor for talking-head Reels, TikToks and YouTube Shorts (15-60 s). Cuts a raw clip the way a top short-form editor does - hook on screen by 1.5 s, open loop, one idea per beat, a visual change every 1.5-3 s, payoff, then a CTA or a seamless loop back to t
Install
npx skills add https://github.com/Viprasol-Tech/claude-video-editing-skills/tree/main/plugins/video-studio/skills/retention-reel
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install viprasol-tech-claude-video-editing-skills@llmmart
git clone https://github.com/Viprasol-Tech/claude-video-editing-skills.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole viprasol-tech/claude-video-editing-skills collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
Retention Reel
Turns one raw talking clip into a short that holds the scroll: hook by 1.5 s, no dead air, payoff, loop or CTA.
- Best for: tips, opinions, stories, announcements, founder clips, one person talking to camera
- Inputs: one talking clip (any orientation; 16:9 gets a face-follow reframe), optional music, optional B-roll folder
- Default skin:
halo-cut· alternativeshook-lab,bold-bounce,depth-type(text behind the head) - Formats: 9:16 1080x1920 master, 4:5 feed, clean (no captions),
captions.srt - Duration target: 15-60 s (sweet spot 20-40 s); genre preset
retention-reel
Before you start
- Needs the video-studio-core skill in the same skills folder.
CORE=video-studio-core,SKILLS= the folder that holds both, so a skin isSKILLS/halo-cut. - First time on a machine:
python CORE/scripts/check_env.pyand fix what it prints. - Read the preset once:
python CORE/scripts/editor.py genre retention-reel. - Work in a new folder next to the footage (
edit-<name>/), never inside a skill folder. - The
--genrepreset sets autocut's pause cut (0.3 s) and run.py respects it; pass--max-gaponly to override.
How a pro cuts this
Structure (percent of the finished length):
| beat | % | what happens |
|---|---|---|
| hook | 0-6 | the strongest line or visual, on screen as text (<= 7 words), lands by 1.5 s |
| open loop / setup | 6-18 | a number, question, contrast or stakes the end will pay off |
| value beats | 18-76 | one idea per beat, each 3-6 s, something changes every 1.5-3 s |
| payoff | 76-92 | close the loop; the line people will share |
| loop-back / CTA | last 2 s | a CTA, or the last line runs straight into the first frame |
- Pacing: ASL 1.5-3 s. Vary it: a 4 s hold on the emotional line, 1 s snaps in a list. Never uniform.
- Cut rules: cut at the end of a thought, in the quietest gap (
--smart-cuts), keep 0.1 s of breath before a word. Jump cuts are allowed here, but alternate punch-in sizes (1.12x / 1.07x) so it doesn't feel like a metronome. Keep the last good take (--retakes). - Never open with "hi guys", "welcome back", "in this video", a logo or > 0.3 s of silence. Start on the hook sentence even if it was said at 0:40 (
story.py reorder). - J/L cuts: rarely needed in a single-camera reel; use an L-cut only when B-roll runs over a statement.
- Transitions: cut. Zoom punch on key words, an occasional whip or flash on a topic change. One transition kind per reel.
- Music/SFX: optional bed 20-24 LU under the voice, ducked. SFX medium density (~10-15 per minute): pop on key words, swish on cuts, impact on the payoff. Drop the music 0.5 s before the payoff line.
- Colour: correct first, then
clean-commercialorpunchyat 0.6-0.8. Moody topics:moody0.5. - Captions: motion.md preset
hormozi(or the skin's own), 1-3 words, <= 17 cps, inside 10-85 % height and clear of the right 12 %,dodgeFace: true. - Client judges: 3-second hold, watch-through %, shares and saves, "does it look native".
- Common mistakes: slow open, every pause removed (breathless), 3+ overlays stacked, captions under the platform buttons, an outro after the payoff, a hook that the video never pays off.
Make the edit
python CORE/scripts/run.py clip.mp4 --style SKILLS/halo-cut --genre retention-reel \
--smart-cuts --retakes --work edit-reel --stop-after plan
# 16:9 source? add --reframe 9:16 (face-follow crop before prep)
# B-roll folder? add --broll broll/ (files named after what they show)
python CORE/scripts/story.py outline --work edit-reel --genre retention-reel # read outline.md: ranked hooks
If the best hook is not the first sentence, build a cold open and re-run the plan on it:
python CORE/scripts/story.py reorder --work edit-reel --order "s5,s1-s4,s6-s9" --genre retention-reel --format 9:16
# read story_flags.md (Frankenbite / back-reference warnings), then:
python CORE/scripts/run.py - --timeline edit-reel/story_timeline.json --style SKILLS/halo-cut --genre retention-reel \
--smart-cuts --work edit-reel2 --stop-after plan
Then review, fix and look:
python CORE/scripts/editor.py review edit-reel/plan.json --genre retention-reel --fix # read review.md
python CORE/scripts/render.py edit-reel/plan.json --frames 0.5,2,6,12,18 # LOOK at contact_sheet.png
python CORE/scripts/render.py edit-reel/plan.json
python CORE/scripts/finish.py edit-reel/plan.json --platform reels --formats 4:5
Refine the plan
- Hook text: rewrite the hook element to <= 7 words that promise the payoff ("I lost $40k doing this"). Start 0.0 s, hold to ~2.4 s.
- Captions:
{"type": "captions", "preset": "hormozi", "keywords": ["40k", "never"], "dodgeFace": true, "cpsMax": 17}. - Zooms:
speaker.zoomson 4-8 key words, start 0.05 s before the word, scale 1.12 then 1.07, never closer than 2.4 s apart. - Numbers: a
counter("from": 0, "to": 40000, "prefix": "$") for every number that matters. - Lists: a
checklistorstepselement withitems[].at= the second each item is said. - Open loop: a
chiporprogress("position": "top") to show there is an end worth waiting for. - Loop ending: trim the last line so its final word runs into the first word of the hook; no endcard. Or a CTA
textin the last 2 s. - Sound:
"sfx": [{"t": 0.0, "kind": "impact", "gain": 0.5}, {"t": 12.3, "kind": "whoosh-short", "gain": 0.45}];"music": {"src": "bed.mp3", "under": 22, "duck": "speech-aware", "fitTo": "beats"}. - Colour:
"color": {"look": "clean-commercial", "intensity": 0.7};"finish": {"grain": 4, "sharpen": 0.3}. - Re-run
editor.py review ... --fixafter every change and look at the frames again. - Face-aimed punch-ins: with
W/track.jsonthe zooms centre on the face instead of the frame centre: afterrun.py ... --stop-after cut(--stop-after timelinefor a--timelinespec) runpython CORE/scripts/track.py analyze W/cut.mp4 --work W, thenmake_plan.py(it copies the track into the plan). Per zoom,"cx"/"cy"(0-1) aim elsewhere;"toFace": falsekeeps centre.
Deliver
python CORE/scripts/deliver.py edit-reel/plan.json --to reels,tiktok,shorts --out-dir edit-reel/deliver
python CORE/scripts/qc.py edit-reel/deliver/final_reels.mp4 --plan edit-reel/plan.json --platform reels --strict
deliver.pywritescaptions.srt/.vtt, a cover thumbnail andpost.md(titles and hashtags from the transcript only).- Clean version: copy
plan.jsontoplan_clean.json, delete thecaptionsand hook elements, set"outputFile": "final_clean.mp4", thenrender.py+finish.pyon it. - Name files
client_project_v001_9x16_captions.mp4; keep every version.
Check before delivering
- Hook on screen and spoken by 1.5 s; first word <= 0.3 s; no greeting.
review.mdhas no HIGH items; contact sheet: nothing over the face, nothing under the right rail, <= 2 overlays plus captions.- Captions match the audio word for word; names and numbers spelled right.
qc.py: -14 LUFS, -1 dBTP, no black or flash frames, 9:16 safe zone passes.- It loops cleanly or ends on the CTA; nothing after the payoff.
- Tell the user: seconds removed, the hook you chose and why, the look, every file path.
Example requests
- "Make this a reel that keeps people watching - hook, captions, zooms, under 40 seconds."
- "Edit my TikTok with bold-bounce, start with the line about losing 40k."
- "Retention edit for Shorts from this 16:9 clip, loop the ending back to the start."
Files (claude-video-editing-skills)
-
SKILL.md 8.1 KB
--- name: retention-reel description: "Professional retention editor for talking-head Reels, TikToks and YouTube Shorts (15-60 s). Cuts a raw clip the way a top short-form editor does - hook on screen by 1.5 s, open loop, one idea per beat, a visual change every 1.5-3 s, payoff, then a CTA or a seamless loop back to the first frame - and delivers 9:16 burned-in, clean and SRT versions. Use when the user says: make this a reel, edit for retention, cut this for TikTok/Shorts, tighten my talking clip, make it go viral, hook + captions + zooms, short-form edit. Pick any Video Studio skin for the look (default halo-cut)." license: MIT --- # Retention Reel Turns one raw talking clip into a short that holds the scroll: hook by 1.5 s, no dead air, payoff, loop or CTA. - **Best for:** tips, opinions, stories, announcements, founder clips, one person talking to camera - **Inputs:** one talking clip (any orientation; 16:9 gets a face-follow reframe), optional music, optional B-roll folder - **Default skin:** `halo-cut` · alternatives `hook-lab`, `bold-bounce`, `depth-type` (text behind the head) - **Formats:** 9:16 1080x1920 master, 4:5 feed, clean (no captions), `captions.srt` - **Duration target:** 15-60 s (sweet spot 20-40 s); genre preset `retention-reel` ## Before you start - Needs the **video-studio-core** skill in the same skills folder. `CORE` = `video-studio-core`, `SKILLS` = the folder that holds both, so a skin is `SKILLS/halo-cut`. - First time on a machine: `python CORE/scripts/check_env.py` and fix what it prints. - Read the preset once: `python CORE/scripts/editor.py genre retention-reel`. - Work in a new folder next to the footage (`edit-<name>/`), never inside a skill folder. - The `--genre` preset sets autocut's pause cut (0.3 s) and run.py respects it; pass `--max-gap` only to override. ## How a pro cuts this **Structure** (percent of the finished length): | beat | % | what happens | | --- | --- | --- | | hook | 0-6 | the strongest line or visual, on screen as text (<= 7 words), lands by 1.5 s | | open loop / setup | 6-18 | a number, question, contrast or stakes the end will pay off | | value beats | 18-76 | one idea per beat, each 3-6 s, something changes every 1.5-3 s | | payoff | 76-92 | close the loop; the line people will share | | loop-back / CTA | last 2 s | a CTA, or the last line runs straight into the first frame | - **Pacing:** ASL 1.5-3 s. Vary it: a 4 s hold on the emotional line, 1 s snaps in a list. Never uniform. - **Cut rules:** cut at the end of a thought, in the quietest gap (`--smart-cuts`), keep 0.1 s of breath before a word. Jump cuts are allowed here, but alternate punch-in sizes (1.12x / 1.07x) so it doesn't feel like a metronome. Keep the last good take (`--retakes`). - **Never open with** "hi guys", "welcome back", "in this video", a logo or > 0.3 s of silence. Start on the hook sentence even if it was said at 0:40 (`story.py reorder`). - **J/L cuts:** rarely needed in a single-camera reel; use an L-cut only when B-roll runs over a statement. - **Transitions:** cut. Zoom punch on key words, an occasional whip or flash on a topic change. One transition kind per reel. - **Music/SFX:** optional bed 20-24 LU under the voice, ducked. SFX medium density (~10-15 per minute): pop on key words, swish on cuts, impact on the payoff. Drop the music 0.5 s before the payoff line. - **Colour:** correct first, then `clean-commercial` or `punchy` at 0.6-0.8. Moody topics: `moody` 0.5. - **Captions:** motion.md preset `hormozi` (or the skin's own), 1-3 words, <= 17 cps, inside 10-85 % height and clear of the right 12 %, `dodgeFace: true`. - **Client judges:** 3-second hold, watch-through %, shares and saves, "does it look native". - **Common mistakes:** slow open, every pause removed (breathless), 3+ overlays stacked, captions under the platform buttons, an outro after the payoff, a hook that the video never pays off. ## Make the edit ```bash python CORE/scripts/run.py clip.mp4 --style SKILLS/halo-cut --genre retention-reel \ --smart-cuts --retakes --work edit-reel --stop-after plan # 16:9 source? add --reframe 9:16 (face-follow crop before prep) # B-roll folder? add --broll broll/ (files named after what they show) python CORE/scripts/story.py outline --work edit-reel --genre retention-reel # read outline.md: ranked hooks ``` If the best hook is not the first sentence, build a cold open and re-run the plan on it: ```bash python CORE/scripts/story.py reorder --work edit-reel --order "s5,s1-s4,s6-s9" --genre retention-reel --format 9:16 # read story_flags.md (Frankenbite / back-reference warnings), then: python CORE/scripts/run.py - --timeline edit-reel/story_timeline.json --style SKILLS/halo-cut --genre retention-reel \ --smart-cuts --work edit-reel2 --stop-after plan ``` Then review, fix and look: ```bash python CORE/scripts/editor.py review edit-reel/plan.json --genre retention-reel --fix # read review.md python CORE/scripts/render.py edit-reel/plan.json --frames 0.5,2,6,12,18 # LOOK at contact_sheet.png python CORE/scripts/render.py edit-reel/plan.json python CORE/scripts/finish.py edit-reel/plan.json --platform reels --formats 4:5 ``` ## Refine the plan - **Hook text:** rewrite the hook element to <= 7 words that promise the payoff ("I lost $40k doing this"). Start 0.0 s, hold to ~2.4 s. - **Captions:** `{"type": "captions", "preset": "hormozi", "keywords": ["40k", "never"], "dodgeFace": true, "cpsMax": 17}`. - **Zooms:** `speaker.zooms` on 4-8 key words, start 0.05 s before the word, scale 1.12 then 1.07, never closer than 2.4 s apart. - **Numbers:** a `counter` (`"from": 0, "to": 40000, "prefix": "$"`) for every number that matters. - **Lists:** a `checklist` or `steps` element with `items[].at` = the second each item is said. - **Open loop:** a `chip` or `progress` (`"position": "top"`) to show there is an end worth waiting for. - **Loop ending:** trim the last line so its final word runs into the first word of the hook; no endcard. Or a CTA `text` in the last 2 s. - **Sound:** `"sfx": [{"t": 0.0, "kind": "impact", "gain": 0.5}, {"t": 12.3, "kind": "whoosh-short", "gain": 0.45}]`; `"music": {"src": "bed.mp3", "under": 22, "duck": "speech-aware", "fitTo": "beats"}`. - **Colour:** `"color": {"look": "clean-commercial", "intensity": 0.7}`; `"finish": {"grain": 4, "sharpen": 0.3}`. - Re-run `editor.py review ... --fix` after every change and look at the frames again. - **Face-aimed punch-ins:** with `W/track.json` the zooms centre on the face instead of the frame centre: after `run.py ... --stop-after cut` (`--stop-after timeline` for a `--timeline` spec) run `python CORE/scripts/track.py analyze W/cut.mp4 --work W`, then `make_plan.py` (it copies the track into the plan). Per zoom, `"cx"`/`"cy"` (0-1) aim elsewhere; `"toFace": false` keeps centre. ## Deliver ```bash python CORE/scripts/deliver.py edit-reel/plan.json --to reels,tiktok,shorts --out-dir edit-reel/deliver python CORE/scripts/qc.py edit-reel/deliver/final_reels.mp4 --plan edit-reel/plan.json --platform reels --strict ``` - `deliver.py` writes `captions.srt/.vtt`, a cover thumbnail and `post.md` (titles and hashtags from the transcript only). - **Clean version:** copy `plan.json` to `plan_clean.json`, delete the `captions` and hook elements, set `"outputFile": "final_clean.mp4"`, then `render.py` + `finish.py` on it. - Name files `client_project_v001_9x16_captions.mp4`; keep every version. ## Check before delivering - Hook on screen and spoken by 1.5 s; first word <= 0.3 s; no greeting. - `review.md` has no HIGH items; contact sheet: nothing over the face, nothing under the right rail, <= 2 overlays plus captions. - Captions match the audio word for word; names and numbers spelled right. - `qc.py`: -14 LUFS, -1 dBTP, no black or flash frames, 9:16 safe zone passes. - It loops cleanly or ends on the CTA; nothing after the payoff. - Tell the user: seconds removed, the hook you chose and why, the look, every file path. ## Example requests - "Make this a reel that keeps people watching - hook, captions, zooms, under 40 seconds." - "Edit my TikTok with bold-bounce, start with the line about losing 40k." - "Retention edit for Shorts from this 16:9 clip, loop the ending back to the start."
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.