short-videos
Produce 90-second universal short-video scripts (HeyGen / Instagram Reels / LinkedIn Video / YouTube Shorts) and standalone LinkedIn posts using native Hermes tools. No Claude Code dependency.
- License 许可
- MIT
- 最近更新
- 2026-11-10
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- short-videos
- license
- MIT
- metadata
- { "tags": [ "video", "linkedin", "social-media", "content", "short-form", "heygen" ], "version": "1.2.0", "platforms": [ "linux" ], "related_skills": [ "medium-story", "technical-writing" ] }
- description
- Produce 90-second universal short-video scripts (HeyGen / Instagram Reels / LinkedIn Video / YouTube Shorts) and standalone LinkedIn posts using native Hermes tools. No Claude Code dependency.
02 资源文件 3
- 其他.gitignore370 B
- 参考文档README.md7.1 KB
- 参考文档VIDEO_FORMAT.md8.3 KB
03 SKILL.md 正文
Short Videos — Hermes-Native Video Script Pipeline
Produce a 90-second universal short-video script and standalone LinkedIn post using Hermes tools directly. No tmux, no Claude Code CLI.
Configuration Variables
| Variable | Default | Description |
|---|---|---|
{{REPO_ROOT}} |
(required) | Path to your Git repository root, e.g. ~/projects/my-content-repo |
{{RP_VIDEO_DIR}} |
{{REPO_ROOT}}/short_videos |
Directory holding numbered video folders under the repo root |
Before running the pipeline, set {{REPO_ROOT}} to your actual content repository path. All file operations reference this variable — never hardcode absolute paths.
Prerequisites
- Hermes Agent (runtime environment)
- Git repo (your content repository — clone it to
{{REPO_ROOT}}) - GitHub / Git remote configured for push
- HeyGen (optional) — scripts are HeyGen-ready but work on any teleprompter platform
- Python 3 for word-count verification
When to Use
- User says "create a short video about..." or "make a reel on..."
- User wants a LinkedIn post that stands alone (not marketing for a video or article)
- User needs a HeyGen script for their digital twin
Outputs Produced
| File | Format | Description |
|---|---|---|
video-script.md |
Markdown (90s, timestamped blocks) | Universal teleprompter script for HeyGen/Reels/Shorts/LinkedIn Video |
linkedin-post.md |
Markdown | Standalone LinkedIn post (no article link, ≤3,000 chars, 5 hashtags) |
quality-report.md |
Markdown | Fact-check, word count, British English, hook strength, character count |
Repo & Conventions
- Repo path:
{{REPO_ROOT}} - Video folders:
{{RP_VIDEO_DIR}}/NN_slug/— numbered sequentially, zero-padded to 2 digits (01, 02, 03...) - Root files:
CLAUDE.md(project instructions),techblogger.md(writer persona),md_to_html.py - Sibling pipeline:
medium-storywrites tounpublished_stories/— this pipeline writes to{{RP_VIDEO_DIR}}/in the same repo
Pipeline Steps
Step 1: Sync the repo
terminal(command="git -C {{REPO_ROOT}} pull --ff-only origin main", timeout=30)
Step 2: Read context files and find next video number
Read these files from the repo root:
CLAUDE.md— project-level instructionstechblogger.md— writer persona
Find the next video number:
terminal(command="ls {{RP_VIDEO_DIR}}/ 2>/dev/null | grep -E '^[0-9]+_' | sed 's/_.*//' | sort -n | tail -1")
If the folder doesn't exist yet, start at 01. Otherwise next = that value + 1, zero-padded to 2 digits.
Step 3: Research the topic
Spawn a research subagent via delegate_task. The research must produce:
- One sharp angle / thesis — the single most provocative or useful truth about this topic
- 3-5 concrete facts or statistics (UK/EU sources preferred; US only where no EU equivalent exists)
- One real-world example or incident that makes the abstract tangible
- Quotable phrasing candidates — short, punchy, tweetable sentences
- Audience relevance — why does this matter right now to infrastructure and security professionals?
Step 4: Create video directory
terminal(command="mkdir -p {{RP_VIDEO_DIR}}/NN_slug/")
Step 5: Run three parallel agents
Use delegate_task with tasks: [...] to spawn all three simultaneously. Pass research context to each.
Agent A — Universal Video Script (video-script.md)
Write a 90-second teleprompter script for the presenter's HeyGen digital twin, designed for:
- Instagram Reels (9:16, vertical, mobile-first, caption-dependent viewers)
- LinkedIn Video (square or 9:16, professional audience, often watched muted)
- YouTube Shorts (9:16, vertical, algorithm-driven discovery)
Script structure — use exactly this format:
TOPIC: [topic name]
SPOKEN WORD COUNT TARGET: 200–230 words
PLATFORMS: Instagram Reels · LinkedIn Video · YouTube Shorts
---
[HOOK] 00:00–00:05
Spoken: ...
On-screen text: ...
Stage direction: ...
B-roll / graphic: ...
[SETUP] 00:05–00:20
Spoken: ...
On-screen text: ...
Stage direction: ...
B-roll / graphic: ...
[INSIGHT 1] 00:20–00:40
Spoken: ...
On-screen text: ...
Stage direction: ...
B-roll / graphic: ...
[INSIGHT 2] 00:40–01:00
Spoken: ...
On-screen text: ...
Stage direction: ...
B-roll / graphic: ...
[PAYOFF] 01:00–01:20
Spoken: ...
On-screen text: ...
Stage direction: ...
B-roll / graphic: ...
[CTA] 01:20–01:30
Spoken: ...
On-screen text: ...
Stage direction: ...
B-roll / graphic: ...
---
CAPTION (universal — use for all three platforms):
[3-5 lines: hook, core insight, 5 hashtags. No "link in bio".]
THUMBNAIL CONCEPT:
[One sentence describing a still frame that works as a cover image across all platforms.]
Script rules:
- Hook lands within the first 3 seconds — bold claim, provocative question, or counterintuitive fact
- Spoken text must sound natural read aloud at steady pace — no bullet points, no headers, no jargon dumps
- On-screen text: max 6 words per frame — reinforce spoken word, don't duplicate verbatim
- Stage directions: brief cues (e.g. "lean forward slightly", "pause for emphasis")
- B-roll/graphic: practical suggestions (e.g. "terminal output animation", "NCSC logo fade in")
- British English throughout
- No mid-sentence dashes — use commas, colons, or semicolons
- No fluffy openers — do not start with "I've been in infrastructure for X years"
- End CTA must be platform-neutral — no "swipe up", no "link in bio". Use "follow for more", "drop a comment", or "save this one"
Word count verification (mandatory after each draft): After writing the script, extract and count the spoken words to verify the total is within 200–230:
python3 -c "
import sys, re
text = open(sys.argv[1]).read()
# Match 'Spoken: <text>' lines in inline-colon format
blocks = re.findall(r'^Spoken:\\s*(.*?)$', text, re.MULTILINE)
total = sum(len(b.split()) for b in blocks)
print(f'Spoken word count: {total}')
if total < 200: print('TOO LOW — expand spoken sections')
elif total > 230: print('TOO HIGH — trim spoken sections')
else: print('Within target range.')
"
Run this after every edit to the spoken text until the count is between 200 and 230. Do not guess — verify.
Agent B — LinkedIn Post (linkedin-post.md)
Write a standalone LinkedIn post — not marketing for a video or article. It stands alone as a piece of insight content.
- Hook line: must stop the scroll in the first sentence
- Body: 3-6 short paragraphs, each punchy and self-contained
- No "link in comments" — there is no link
- No "watch my video" language — this post exists independently
- 3,000 character hard limit (LinkedIn cap) — count carefully
- End with a question or provocation that invites comments
- Exactly 5 hashtags at the end
- British English, no mid-sentence dashes
- Do not open with "I've been in infrastructure for X years" or any years-of-experience framing
Agent C — Quality Check (quality-report.md)
Review both video-script.md and linkedin-post.md once drafted, producing a report covering:
- Fact accuracy — verify each statistic and claim; flag anything unverifiable or misleading
- Spoken word count — count spoken sections of video script; flag if outside 200-230 words
- Timing sense-check — does script feel paced correctly for 90 seconds at natural speech rate?
- British English — flag any Americanisms
- Dash check — flag any mid-sentence em dashes or en dashes used as parenthetical breaks
- LinkedIn character count — report exact count and flag if over 3,000
- Hook strength — rate the video hook and LinkedIn hook on 1-5 scale with reasoning
- CTA check — confirm video CTA contains no platform-specific mechanics
If issues found, list as actionable items. Do not rewrite the files — report only.
Step 6: Commit and push
terminal(command="cd {{REPO_ROOT}} && git add {{RP_VIDEO_DIR}}/NN_slug/ && git commit -m 'Add short video NN: [topic summary]' && git push", timeout=30)
Always push. Never leave new video content uncommitted.
Writer Voice & Style Guide
This section is derived from analysing a practitioner's published work across multiple platforms. The patterns below describe what the voice actually sounds like — not a theoretical style guide.
Core Voice
- Practitioner, not pundit. Speaks from experience managing real infrastructure at scale.
- British English always. -ise not -ize, colour, favour, behaviour, defence, whilst, amongst, licence, harbour, realise, analyse.
- Dry, understated wit. Surfaces naturally when the situation is absurd. "The catch, in this case, is that there is not one." / "I collect them the way people collect gym memberships: with great optimism and almost zero follow-through."
- Leads with impact, then explains mechanism. The human/operational consequence first, the technical explanation second.
Short Video Script Voice (90 seconds, spoken)
Because the short video format compresses everything, the voice translates differently:
- Hook lands by reframing or surprising. Not a question. Not a clickbait opener. A short, sharp statement of the unexpected truth.
- Sentences are shorter than in articles but maintain the same rhythm. 15-25 words spoken per sentence. Still precise. Still British English. Still no filler.
- "Here is the thing" / "Here is what matters" are still banned. Say the thing. Don't announce you're about to say it.
- End CTA is direct and platform-neutral: "Follow for more." / "Drop a comment." / "Save this one." Never "swipe up", never "link in bio."
LinkedIn Post Voice (standalone)
- Opens like telling a colleague what happened. "Ran the NCSC exercise with Rui this week." / "Microsoft Edge decrypts every single saved password the moment it launches."
- No emoji-headlines. No "🛑 What would you do if...?" — that's ad copy, not the writer's voice.
- 3-6 short paragraphs, each punchy. 3,000 character hard limit. Every sentence earns its place.
- No "with over 20 years of experience" framing — authority comes from the content, not the opener.
- Ends with a question or provocation that invites comments.
- Exactly 5 hashtags.
What NEVER Appears
- ❌ Mid-sentence dashes (em dashes, en dashes) — use commas or parentheses
- ❌ AI transition phrases: "Here's what matters", "What this means is", "Here's the thing"
- ❌ Corporate jargon: "data estate", "operational reality", "measured in inconvenience"
- ❌ Fluffy openers: "In today's digital landscape"
- ❌ Credentials-first: "With over 20 years of experience..."
- ❌ Ad-copy hooks: "Not hypothetically. In the next 60 seconds."
- ❌ Paired oppositions (language-model parallelism)
- ❌ American spellings — ever
- ❌ "It is important to note that" — just note it
- ❌ "Furthermore", "Moreover", "Additionally" — use "And"
The Test
Read it aloud. If it sounds like a LinkedIn post from someone trying to sound authoritative, cut it. If it sounds like someone who just got off a call with a vendor who overpromised and underdelivered — that is the voice.
Pre-Publish Checklist
Before saving any script or LinkedIn post, run this check:
| Pattern | Fix |
|---|---|
| "delve", "tapestry", "realm", "leverage" | Replace with plain English |
| "in order to" | Replace with "to" |
| "it is worth [verb]ing" | Replace with "Let me [verb]" |
| "it is important to note that" | Delete |
| "Furthermore," / "Moreover," | Replace with "And" or nothing |
| "Not only... but also" | Simplify |
| "By [verb]ing..." (sentence opener) | Rephrase |
| "One of the most [adjective]" | Be specific or delete |
| "In today's [X]" / "In an era of" | Delete entirely |
| Contractions missing | Add don't/won't/it's/can't |
Run this checklist before returning. If any pattern appears, rewrite.
- Authority comes from tone and insight, not from stating credentials
Pitfalls
- Research agents may fabricate statistics. The quality check agent catches these — trust but verify. The quality report is a genuine check, not a rubber stamp.
- LinkedIn post is standalone. No "link in bio/description" CTA. The quality agent enforces this.
- Video script CTA must be platform-neutral. No "swipe up", no "link in bio". Use "follow for more", "drop a comment", "save this one".
- Spoken word count target is 200-230 words. Shorter scripts feel rushed or sparse; longer scripts exceed 90 seconds at natural speech rate.
- Hook must land within 3 seconds. Vertical video viewers decide whether to watch in the first 1-2 seconds. The hook is the single most important element.
- Git authentication: If
git pushfails, report the error. The files exist locally — don't lose them. - Delegate_task context limits: For the video script agent especially, the output format is very specific. Include the full script template in the context to ensure correct structure.
04 同仓库更多 Skills
writing-dna-skill
从至少 20 篇完整文章中蒸馏可复用的写作 DNA,分析语言、文章结构、选题逻辑、素材策略、认知框架和视觉风格,并生成 Writing-DNA.md;按该 DNA 写作时会读取全部蒸馏产物和 5 篇相关原文。用于中英文作者、账号、品牌或出版物的风格分析与一致性写作。 Distill reusable Writing DNA from at least 20 complete articles for Chinese or English authors, publications, brands, and accounts; use for language, structure, topic logic, source strategy, cognitive-frame, and visual-style analysis.
ub-quality-gates
Uncle Bob 编程纪律与确定性质量门工具链。写代码或改代码时使用:先让它能跑,再让它正确,最后守纪律。 当用户要求 /ub-mode、需要代码质量门禁、CRAP 评分、重复代码检测、变异测试、 或抱怨代码难懂/难测/有重复时触发。覆盖 Clojure/Java/Go/TypeScript/Rust/Python。
tweet-composer
Score and optimize tweets based on X's real open-source ranking algorithm. Analyzes draft tweets against the actual ranking code — not generic tips. Use when: composing tweets, optimizing drafts for reach, planning threads, analyzing why a tweet performed well/poorly, or asking for posting strategy advice.
short-video-scripter
Use when the user asks to "script this short video", "write a TikTok / Reels / Shorts script", "给这条抖音或视频号视频写脚本", or "fix the hook — viewers drop off in the first seconds"; produces timestamped beat-sheet scripts on the retention-gate model (0-2s hook / 2-5s confirmation / 5-15s payoff / loop-or-CTA) with per-scene script lines, on-screen text for muted viewing, asset keywords, and 9:16 format-param rows per platform norm card for TikTok, Reels, Shorts, 抖音, and 视频号 — plus 2-3 hook options from named hook families and an AI-content disclosure line by default on realistic synthetic media. Spec-only: rendering, TTS, and publishing stay with the user's own tools — no pipelines, no upload automation. Not for creator video briefs — use brief-generator; long-form video SEO belongs to content-writer. 短视频脚本/抖音分镜/开头钩子/竖屏9:16
remotion-video-master
Remotion 视频编程大师 — 用 React 组件以编程方式创建 MP4 视频。支持动画、字幕、音乐可视化、 AI 配音(TTS)与场景自动配时。触发词:用代码做视频、编程视频、React 视频、Remotion、 /remotion-video。适用场景:程序化视频(批量生成、数据驱动如年度总结、音乐可视化、自动字幕)、 教程讲解视频(技术概念可视化如 CNN 与算法、分层递进讲解、AI 配音教程)、 3D 视频(产品展示与模型动画、卡通角色讲解、3D 数据可视化、Logo 动画)。
medium-story
Full Medium article pipeline using native Hermes tools. Research → write → 4 parallel output agents → HTML conversion → git commit/push. Includes pre-flight infrastructure verification step, revisor-methodology fact-checking, and an anti-AI-slop writing pass (banned vocabulary, structural variety, accuracy rules).