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

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

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-story writes to unpublished_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 instructions
• techblogger.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:

1. Fact accuracy — verify each statistic and claim; flag anything unverifiable or misleading
2. Spoken word count — count spoken sections of video script; flag if outside 200-230 words
3. Timing sense-check — does script feel paced correctly for 90 seconds at natural speech rate?
4. British English — flag any Americanisms
5. Dash check — flag any mid-sentence em dashes or en dashes used as parenthetical breaks
6. LinkedIn character count — report exact count and flag if over 3,000
7. Hook strength — rate the video hook and LinkedIn hook on 1-5 scale with reasoning
8. 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:

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 push fails, 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.