demo-reel
demo-reel turns your repo, website, or screen recording into viral-ready reels for TikTok, Instagram Reels, YouTube Shorts, LinkedIn, and X. Each run makes 3 hook variants in 3 native formats and verifies every render. After you post, it picks the winning hook from real views.
$ npx skills add https://github.com/shashankswe2020-ux/demo-reel --skill demo-reel
Star on GitHub → See real results
Then run /demo-reel in Claude Code, Codex CLI, GitHub Copilot, Cursor, or any agent that supports skills.
Every step ends in a gate. When a render fails, the agent fixes the cause, re-renders, and re-checks. It never lowers the bar.
Pulls the entry → action → result flow, exact colors and fonts, a lexicon of product words, and claims with verbatim source quotes.
Hooks across at least 4 types are ranked by a published formula. The top 3 become A/B variants.
reel-plan.json must pass 29 checks: reading time, pattern interrupts, safe zones, CTA, and share-copy limits.
One deterministic composition, laid out separately for 9:16, 1:1, and 16:9, then finished to −14 LUFS with a poster baked into frame 0.
22 pixel and audio gates per render: first motion, black frames, frozen stretches, true peak, and caption readability.
Post all three variants. A Bayesian P(best) over hook rate, completion, or shares decides what ships next.
Each gate is a threshold that FFmpeg or the linter measures, and every one is listed in metrics.md. A render ships only with zero blockers and a Viral Readiness Score of at least 85.
Both kits are committed with their plans and QA reports, so every number here can be re-checked.
The first gate run scored 91.7 and caught 4 real defects, including 2 toolkit bugs (full-range color and 0 dBTP clipping). After fixes: VRS 100 on 9/9 renders, −14.4 LUFS, first motion at 0.07 s. QA report
10 hooks across 10 types and 4 claims verified against source, including "66 models with evidence attached". VRS 100 on 9/9 renders, −14.0 LUFS, 7–8 cuts per reel. QA report
Each reel shows one headline feature, built from the repo's own site, screenshots, and demo footage in its brand fonts and colors. All 36 renders (12 repos × 3 formats) passed the 51 gates at VRS 100. Click a preview to watch with sound. These are independent demos, not affiliated with or endorsed by the projects.
16:9 landscape: X, YouTube, websites
Describe it, get a diagram
Line-level review comments
Hire the team
Every lesson ships an artifact
1:1 square: LinkedIn and X feeds
Plans before it builds
Retain, then recall
Ask like a scientist
Clone a voice, then speak
9:16 vertical: TikTok, Instagram Reels, YouTube Shorts
The browser has one
Action first. Steps numbered.
A topic becomes a classroom
ChatGPT Web in Codex
Install the skill, open your repo in an agent that supports skills, and run /demo-reel. It reads the code, writes the hooks, renders the videos, and checks them.
Vertical 9:16 (1080×1920) for TikTok, Instagram Reels, and YouTube Shorts. Square 1:1 for LinkedIn and X feeds. Landscape 16:9 for X, YouTube, and your website. Each render comes with SRT captions and a poster image.
Any agent that loads Agent Skills: Claude Code, OpenAI Codex CLI, GitHub Copilot, Cursor, Gemini CLI, and opencode.
No. The toolkit is stdlib-only Python plus FFmpeg, and rendering runs on your machine.
No. Every number on screen must quote the project's own source verbatim, and the linter checks each quote before anything renders.