Skill · Workflow
short-form-video-pack
Researches what short-video formats and hooks work in your niche, then writes five reviewed concepts with Veo, Sora or Kling shot prompts. Use when asked to run the "Short-form video pack from what's working now" workflow.
The skill itself. Agents read the name and description first and load the rest only when a task matches.
short-form-video-pack/SKILL.md
---
name: short-form-video-pack
description: Researches what short-video formats and hooks work in your niche, then writes five reviewed concepts with Veo, Sora or Kling shot prompts. Use when asked to run the "Short-form video pack from what's working now" workflow.
---
# Short-form video pack from what's working now
Researches what short-video formats and hooks work in your niche, then writes five reviewed concepts with Veo, Sora or Kling shot prompts.
## Inputs
Ask for any of these that the user hasn't given:
- `niche` (e.g. "home espresso for beginners")
- `topic` (e.g. "our $149 manual milk frother (the Froth One)")
- `platform` (e.g. "TikTok")
- `audience` (e.g. "25-40 year olds who just bought their first espresso machine")
- `length` (e.g. "30-40 seconds")
- `brand_rules` (e.g. "No competitor names. No health claims about caffeine. Founder can film talking-head clips in a home kitchen; no paid actors.")
- `format_preference` (e.g. "auto")
## Steps
Run the steps in order. Refer to saved results by name. Don't skip a step that says STOP.
1. **Manual start** — Start: {{niche}}, {{topic}}, {{platform}}, {{audience}}, {{length}}, {{brand_rules}}; {{format_preference}}: 'auto', or talking head, faceless b-roll, ASMR, product demo, AI avatar.
2. **Research what's working now** — Find short videos about {{niche}} that got traction on {{platform}} in the last 90 days, via creator-economy articles, trend pages, Reddit and newsletters. You can't watch videos: use titles, captions, transcripts and write-ups, and say which.
Return: 8-12 working patterns (format, hook verbatim or marked 'paraphrased', length, why it works, URL and date); 3 overdone patterns; hashtags seen in 2+ sources; {{topic}}-specific launches, complaints or questions.
Every row needs a URL. View counts only if a source states them. Under 5 sources from the last 90 days: say so first and date the older ones. Save the result as `research`.
3. **Pick the format** — Use {{format_preference}} if it names a route; else the format most common in recent {{research}} patterns that suits {{topic}} and fits {{brand_rules}}. Pick one route: Save the result as `format`.
- **Talking head**: Talking-head playbook ({{length}}, {{platform}}, {{topic}}, from {{research}}), first line 'Format: Talking head': timestamped beats (hook 0-2s, setup, 2-3 payoffs, CTA); a cut every 2-4s; b-roll about every 8s; on-screen text for muted viewing; phone at eye height, window light, clip-on mic; 3 research hooks with URLs; 2 common mistakes.
- **Faceless b-roll**: Faceless b-roll playbook ({{length}}, {{platform}}, {{topic}}, from {{research}}), first line 'Format: Faceless b-roll': timestamped beats; 1.5-3s clips; voiceover about 2.5 words/s; on-screen text that carries the story on mute; clip types per beat; 3 research hooks with URLs; 2 common mistakes.
- **ASMR or satisfying**: ASMR playbook ({{length}}, {{platform}}, {{topic}}, from {{research}}), first line 'Format: ASMR': timestamped beats with sound as the hook; the sound per beat (pour, tap, foam, click); slow macro close-ups; little or no voice; last frame matches the first (loop); 3 research hooks with URLs; 2 common mistakes.
- **Product demo**: Product demo playbook ({{length}}, {{platform}}, {{topic}}, from {{research}}), first line 'Format: Product demo': timestamped beats (product in hand in second 1, problem, use, visible result, CTA); real scale and use; which claims need proof on screen; disclosure if paid or gifted; 3 research hooks with URLs; 2 common mistakes.
- **AI avatar**: AI avatar playbook ({{length}}, {{platform}}, {{topic}}, from {{research}}), first line 'Format: AI avatar': timestamped beats; lip-sync rules (sentences under 15 words, numbers written out, no tongue-twisters); b-roll cutaways every 5-8s; the platform's AI-content label; no real person's face or voice without written permission; 3 research hooks with URLs; 2 common mistakes.
4. **Five concepts at once** — Launch these as parallel subagents in one message, then wait for all of them. Write one concept for {{topic}} per the {{format}} playbook, for {{audience}} on {{platform}}, at {{length}}, within {{brand_rules}}, borrowing a {{research}} pattern if it fits.
Return: title (5 words max); hook line (12 words max) and first-2s on-screen text (6 words max); timestamped script for {{length}} at about 2.5 words/s; on-screen text per beat; retention beats (timestamps of reveals or cuts); one closing CTA; borrowed pattern and URL.
No facts about {{topic}} beyond the inputs or cited sources; mark what to confirm [confirm]. Save the result as `concepts`.
- **Concept 1: myth or mistake**: Name a common belief or mistake in {{niche}}, then correct it.
- **Concept 2: how-to**: Teach one result in 3 steps the viewer can copy today.
- **Concept 3: story or before/after**: A story or before/after shown on screen.
- **Concept 4: test or comparison**: Compare 2-3 options with a verdict; describe them generically if brand rules forbid competitor names.
- **Concept 5: trend remix**: Adapt the strongest recent pattern's structure without copying its lines.
5. **Write shot prompts** — Load the `prompts-gpt-video-prompt` skill, then: For each concept in {{concepts}}, repeat its title, hook, script, on-screen text and CTA, then a shot list per {{format}}, so this is the whole pack.
Per shot: number, timestamps, duration (8s max: one clip in models such as Veo; check yours), filmed or generated. Generated shots get a prompt: subject, action, setting, camera (framing, lens, movement), lighting, style, sound, 9:16, with the same subject description across a concept. No real people's likeness, logos or copyrighted characters. Durations total the script within 3s. Save the result as `pack`.
6. **Critic review** — Review loop (max 2 passes): Score each concept 1-5 per point; pass only if all score 4+.
1. Hook: payoff, tension or surprise in 2s, 12 words max; 'In this video', 'Hey guys' or a logo intro scores 2 or less.
2. Retention: a new beat or visual every 5s or sooner; payoff before the CTA.
3. {{platform}} fit: within {{length}}, text readable in time (about 3 words/s), works on mute, 9:16.
4. Policy: health, money or legal claims only with [confirm]; no real-person likeness or voice, copyrighted music or characters; AI label on generated shots; paid or gifted products disclosed.
5. Shots: 8s max with camera, motion and duration; total within 3s of the script.
Feedback: concept, point, fix; revise only what failed. Save the result as `pack`.
7. **Captions and hashtags** — For each concept in {{pack}}, write {{platform}} copy: a caption opening on the hook (under 150 characters before the fold, 300 total), 3-5 hashtags only from {{research}} plus up to 2 broad {{niche}} ones (call none trending), a pinned comment inviting replies, cover text (4 words max). End with a shoot order, easiest first. Save the result as `captions`.
8. **Pick what to shoot** — STOP and ask the user to approve before continuing: Review concepts, shots, captions and every [confirm]; note which concepts you'll shoot. Approve to export; reject to stop.
9. **Save the pack** — Finish: Return {{pack}} and {{captions}} as one ready-to-shoot document.
## Done when
Every step has run, each approval was given, and you've reported what was produced and where it was saved.