I Automated a YouTube Shorts History Channel — and Here’s What the Data Taught Me
Most evenings — and every spare moment while my younger son works through basketball practice— I’m hunched over a beat-up Chromebook tethered to my phone, welding together a garage-based techno-art studio. A $200 Orange Pi 5 hums inside the network cupboard at home, running the n8n workflows I prototype courtside between free throws. Six AI personas — writer, critic, art director, data wonk, producer, and tech wrangler — take shifts cranking out one-minute history for YouTube Shorts while the dog and I patrol the neighborhood. No grand strategy here; I simply poured two decades of analytics and process-improvement obsession into a curiosity project and let it sprawl. What follows is the wiring diagram — and how those sideline hacks turned idle fascination into a surprisingly sharp crash course on frontier LLMs and agentic AI.
- “Promptlets” keep prompts agile:
Breaking large prompts into Lego-sized promptlet blocks makes it painless to tweak just the hook, the call to action, or the tone. Teams can swap pieces without rewriting (or re-testing) the entire chain. - Use Markdown to shape the conversation
Give each promptlet an#header, then rely on:##for H2 headers and###for H3 headers; Bold with ** for must-follow rules; Bullet lists for step-by-step instructions;Code back-tickswith ```character for variables, file paths, or literal text the model must not change; Numbered lists when order matters; Horizontal rules (---) to separate distinct phases; Clear structure = fewer hallucinations and easier version control. **///**inline comments protect the model from noise:
Anything after a triple-slash is stripped by my n8n workflow before the request reaches the LLM, so you can leave editorial notes in the prompt file without confusing the model or leaking instructions into the output.- Lean Six Sigma for your prompts is a thing: Map the whole prompt flow, hunt down duplicate or clashing rules, dump the waste, and surface just 3–5 guiding principles-exactly how large companies uses Lean Six Sigma to fix bloated SOPs.
- Treat models like teammates, not oracles:
Each LLM gets a clear role (writer, art director, analyst) and its decisions are logged, reviewed and scored, just like human staff. This transparency turns “black-box magic” into measurable performance you can coach — or replace. - YouTube Shorts first 4seconds is life-or-death:
Viewer retention curves fall off a cliff in the first 10 seconds. If the opening shot isn’t irresistible, the algorithm never lets the rest of the video see daylight. First-scene testing beats every other optimization.
Note: The robots you see in the images of this article are creative rendering of InMoov humanoids to illustrate the spirit of the team of GenAI agents — fully printable at home with off-the-shelf servos and a spool of PLA. Their outfits hint at each agent’s role with clear stereotype, but beneath the flair they run the same code powering my garage studio. Want your own AI coworker? Head to inmoov.fr and start printing.
Check out the YouTube Channel at https://www.youtube.com/@AIIllustratedHistory/shorts
1. Why YouTube Shorts?
Shorts are YouTube’s bustling bazaar: seventy‑plus billion daily views of vertical minute-long clips. That makes them catnip for reach-but also a coliseum where attention spans die in under two seconds. For me, Shorts tick three boxes:
- Tight feedback loop. A clip either catches fire or sputters out within 48 hours, giving hard‑numbers feedback on hooks, art, and pacing.
- Automatable canvas. The rigid format (vertical, no B‑roll, one audio track) is perfect for AI‑generated assets. A clear spec means fewer edge cases for the bots to fumble.
- Asymmetric upside. One great Short can rack up more views in a day than a ten‑minute traditional video does in a month-if you stick the landing.
That upside justifies the experimentation. My worst‑performing clip stalled at 0 views (YouTube Shorts Feed Algorithm didn’t like some of my keywords); my best cleared 4.8k in 24 hours with no ad spend. The chasm between those two outcomes is what this article is about.
This whole channel is my techno-artistic pet project. Think of it as the moment a Stone Age painter first traded finger smears for a bristly brush: tradition gasped, then evolved. Today’s brush is AI, and if it shocks some creators, that’s just history repeating itself — only faster.
2. The Garage Tech Stack
I run the entire pipeline on an Orange Pi 5-an $200 single‑board computer that sips five watts at idle and maxes out under 20. Every component is chosen for a balance between performance and flexibility. Here’s the abridged bill of materials:
Hardware: Orange Pi 5 Plus
24/7 uptime, 32 GB of memory, 1 TB of SSD storage, running Ubuntu.
Orchestration: n8n (self‑hosted) “n-odematio-n”
Visual workflow, each action is a node that we drag’n’drop, JSON in/out, zero vendor lock‑in.
LLM Gateway: OpenRouter.ai
Lets me hot‑swap GPT‑4.1, Claude Sonnet 4, Gemini 2.5 pro without code changes. (picked randomly based on status)
Audio: gpt‑4o-mini-tts
Flexible voices with built‑in “how to speak” instructions.
Images: GPT‑4o Images, Google Imagen, or Leonardo (random pick for each batch)
Ongoing A/B/C test of style, speed, and cost.
Video Engine API: MoviePy + OpenCV
Alternate zoom-in and zoom-out, smart zoom via saliency POI detection, and word‑perfect caption timing.
Storage: Github + Google Drive + BigQuery
Houses prompts, model selections, scores, and YouTube analytics for post‑mortems.
If the Orange Pi halts, a watchdog reboots it and a docker compose up brings the whole factory back in minutes-no human lamenting lost configs.
3. Meet the AI Crew
At 6 am each day, and every 3h until 6 pm, I roll a three‑sided die-GPT‑4.1, Claude Sonnet 4, or Gemini 2.5 Pro-and the winner becomes the sole brain for that entire batch (yet another live A/B/C test). With the model locked in, I treat the pipeline as a five‑person virtual studio where each “employee” is just a prompt‑driven persona:
Zoom image will be displayed
Alpha, the Studio Exec
Alpha —the Studio Exec (ChatGPT web, OpenAI Codex). Sets direction and policies, co‑authors “Promptlets” (GenAI prompt sections), and then steps aside when the team is hard at work.
Zoom image will be displayed
Beta, the Creator
Beta — the Creator (frontier LLM). Crafts a ~100‑word voice-over through four to six‑scene with voice guidance on how to speak, as well as the scene illustrations’ prompts, all the YouTube metadata (title, description, filename, hashtags, etc.) and identify optimal background music from a predefined list.
Zoom image will be displayed
Gamma, the Critic
Gamma — the Critic (frontier LLM). Scores the voice-over and illustration prompts on hook/appeal strength, clarity, and compliance, then suggests surgical improvements.
Zoom image will be displayed
Delta, the Producer
Delta — the Producer (n8n orchestrating GenAI API calls). Generates final audio and high‑res illustrations, storing all in a Google Drive folder.
Zoom image will be displayed
Omega, the Techy
Omega — the Techy (Video Editing API in Python). Download all the assets from Google Drive, and animates stills, runs OpenCV‑based saliency zoom on POIs, leverages on‑device STT to align word‑perfect subtitles, muxes everything into an MP4, and triggers the YouTube upload.
Each role writes logs to Google Drive and BigQuery, giving me a timestamped audit trail of who decided what-a lifesaver when debugging a factual flub or style drift.
4 Prompt Engineering Best Practices
4.1 Promptlets: LEGO Bricks for Prompts
Early on I learned that a 2000‑word mega‑prompt is a maintenance nightmare. Enter Promptlets: small, purpose‑scoped blocks tagged with #Tone, #Hook, #Length, #OutputFormat, and so on.
I also invented prompt comment. In coding, each function as a detailed documentation in the code itself, and key statements are preceded by in-comment. I needed context in-line comment in the prompt for Alpha that is managing the ecosystem. So, I started having /// for prompt in-line comment. Everything after a /// is getting stripped before reaching the underlying AI agent (Beta and Gamma).
Need to switch from “cinematic awe” to “snarky sidekick” narration? Open Codex, describe what you are after, get it to think about the problem and propose changes that you review and approve through a regular Github Pull-Request process.
4.2 Lean Six Sigma for Prompts — Elevate Rules to Principles
When a prompt accretes dozens of tactical “thou-shalt” rules, contradictions sneak in and the model stalls. Borrow a page from manufacturing: run a quick DMAIC cycle (Define, Measure, Analyze, Improve, and Control). Define the CTQs-clarity, policy compliance, creative head-room. Measure where breakdowns happen (hallucinations, policy flags, endless revisions). Analyze the text: highlight steps that add no value or outright collide. Improve by merging overlapping rules into a handful of plain-English principles (“one voice, one JSON object, no moral wrap-up”). Control it with version-controlled promptlets and a simple A/B harness so future tweaks stay honest. Two hours of white-boarding can cut a 2 000-word monster prompt down to a lean sheet you can explain at stand-up.
Why it works — Google’s privacy overhaul. In 2012 Google ditched 60-plus product-specific privacy documents and folded them into one policy that was 85 % shorter and far easier to audit. Regulators who had begged for “simpler, clearer rules” suddenly found them, and internal compliance actually improved because engineers could recite the principles instead of hunting through sixty PDFs.
Treating frontier models like human employees is not wrong (it’s not exact, but not wrong). So we can apply the same playbook here: fewer, clearer prompt rules mean your LLM (and your future self at 2 a.m.) will follow them every time. Trust your employees to be smart and able to adjust to changing environment if you provided the right mission, clear objectives and a simple set of common guardrails & guiding principles.
Best-practice cheat sheet
- Start the DMAIC review every time a prompt tops ~800 words.
- Track defect metrics (policy violations, e.g. 3rd arms for a person, manual patches/reruns) so you can see the trend line.
- Store the new “high-level principles” in one
#Governancepromptlet that sits at the top of every chain—everything else can stay modular. Then make sure each promptlet describe more the successful outcome more than the precise run book.
5. A Day in the Life of a Video
Let’s pick a recent video that got 939 views and 14 likes which flow was automatically triggered today, on June 13th 6 am — check it out at www.youtube.com/shorts/-nvgebNGh_g
A. Check Status
First, I check the status of OpenAI (ChatGPT), Anthropic (Claude), OpenRouter.ai, Microsoft Azure and Google Cloud. Based on who is up and running, I define the list of LLM models I can consider, and pick on randomly. Here, openai/gpt-4.1 was the lucky winner.
B. Get Previous Videos
I check on YouTube if I had any other videos for #June13 — and there were two, ran the previous night in California (but still June 13 UTC). One was about the anniversary of Pioneer 10 leaving our solar system on June 13 1983. Another was about the first play of Tempest from Shakespeare on June 13 1611. We will be excluding them from the list of candidates.
C. List Historical Events
Here we start our first AI Agent to list the candidates. We collect the latest prompt from our private Github repository, remove any /// in-line comment for Alpha, and start the AI Agent. An AI Agent is a system with, for example, GPT 4o as its brain, and who will iterate on how to solve a problem through multiple steps. It has access to tools to help augment its knowledge — here Wikipedia, Internet Search, Get Internet Page. This first AI Agent researched through its tools and identify the top 10 candidate events based on its deep-research of historical events that occurred on this day. Alpha had guided in the prompt to veto anything with potential policy baggage (e.g., modern warfare, medical claims). Here, the AI Agent, powered by GPT-4.1 penciled down:
* June 13, 1373 – English mystic Julian of Norwich reports experiencing her Revelations of Divine Love.
* June 13, 1525 – Martin Luther marries Katharina von Bora, setting a precedent for clerical marriage in Protestantism.
* June 13, 1763 – The Royal Proclamation restricting settlement west of the Appalachian Mountains is issued to American colonists by the British.
* June 13, 1789 – Mrs. Alexander Hamilton hosts the first political society for women in the United States in New York City.
* June 13, 1842 – Queen Victoria becomes the first British monarch to travel by train.
* June 13, 1865 – Irish poet and Nobel laureate W. B. Yeats (birth), Dublin, Ireland.
* June 13, 1886 – King Ludwig II of Bavaria (death), famous builder of castles such as Neuschwanstein.
* June 13, 1910 – Mary White (death), inspiration for the famous eulogy by William Allen White.
* June 13, 1920 – The U.S. Post Office Department rules that children may not be sent by parcel post.
* June 13, 1927 – Aviation pioneer Charles Lindbergh receives a hero’s welcome in New York City following his solo transatlantic flight.
* June 13, 1934 – The first U.S. Federal Credit Union is established in North Carolina.
* June 13, 1944 – Germany launches the first V-1 flying bomb attack on London.
* June 13, 1966 – The U.S. Supreme Court rules in Miranda v. Arizona, leading to the "Miranda rights."
* June 13, 1980 – The United Nations Security Council holds its first meeting in Africa, in Addis Ababa, Ethiopia.
* June 13, 1986 – Sir Ben Johnson sets a new world record for the 100-meter sprint at the Canadian Track & Field Championships.
* June 13, 1997 – Timothy McVeigh is sentenced to death for the Oklahoma City bombing.
* June 13, 2005 – Michael Jackson is acquitted of all charges in his child molestation trial
D. Existing Playlists
I collect from YouTube the latest playlists from my Channel. The Metadata AI Agent down the road will be identifying the optimal 1 to 3 playlists to add that video to.
E. Pick a Story
Now we call an n8n workflow where we compartmented the full AI agent flow to analyze the candidates and rank them by perceived virality for modern viewer and emotional stakes. Adding a detailed historical research, including why modern YouTube audience would relate to the story. Here is the detailed analysis from the AI agent:
On this day June 13, 1842 – Queen Victoria becomes the first British monarch
to travel by train.
Why it mattered then: A reigning queen riding a steam train shattered royal
tradition and embraced the shock of new technology.
Story Summary Paragraphs:
1. Britain’s Queen Victoria stunned the world by climbing aboard a rumbling,
smoke-belching steam locomotive—something no monarch had ever dared.
2. Royal advisors fretted about danger and decorum, but Victoria swept past
their doubts, taking her seat and waving as the train lurched into motion.
3. That day, millions saw their queen defy centuries of tradition, turning
the train—once feared as unsafe—into the ultimate symbol of progress.
Outcome: Victoria’s bold ride helped make rail travel trustworthy and
fashionable across the world.
Why it echoes today: When leaders publicly embrace new technology, it
inspires others to dream bigger and trust in progress—just like Victoria’s
train ride did for a whole generation.
Why this resonates with 18–34 U.S. audience: Seeing a young queen defy
expectations to ride cutting-edge tech makes her relatable to anyone who’s
ever tried something new against the odds. It’s the ultimate “main character
energy”—turning a simple train trip into a viral moment that changed society
forever.
Concrete takeaway: Queen Victoria’s first train ride made steam trains wildly
popular and kicked off an era where royalty and regular people alike traveled
side by side.
Big Fifth Grader So What: When a queen tried a new invention, everyone wanted
to try it too. One brave choice can change what people think is possible.
F. Create the voice-over
Now we are ready to call a new sub-workflow dedicated to the voice-over creation. This workflow has two AI Agents: Beta and Gamma. Beta will create a proposal for the voice-over. Gamma will assess and score the work and provide feedback on who to improve it. The two of them work until the score reach 95+ (or ended their 5th iteration). Here is the first output of Beta — in squared brackets [] are the speaking instruction:
Steam blasts out as a royal carriage crashes onto the tracks.
[Kick off with a wild, high-energy burst—like a hype video of a celebrity
doing the impossible]
On this day, June 13, 1842, Queen Victoria boards a thunderous train for
the first time ever.
[Ground the timeline with a breathless, wide-eyed awe—picture standing at
the edge of the platform]
Royal advisors? Losing it. Crowds? Glued to every move as Victoria grabs
her seat.
[Lean in, like sharing forbidden royal tea—delight in the rebellion]
The train jolts. Victoria waves. Her silk dress whips in the wind, steam
swirling everywhere.
[Ramp up the stakes—like a rollercoaster crest, then tip into excitement]
Imagine if Taylor Swift jumped on the first Hyperloop. It was that big.
[Let irony land with a sly, knowing grin—drop a pop-culture beat]
From that moment on, nobody saw trains-or queens-the same way again.
Would you have taken that ride?
[End on awestruck wonder—like the afterglow of a viral moment that lingers] Script Draft (Beta’s Write Voice-Over). Beta returns JSON: title, four‑scene script ≤90 words, SSML cues, image prompts, and metadata.
Gamma’s score was 60 out of 100, with the following recommendations:
suggestion: Open with a concrete, cinematic image: 'Steam explodes as a
royal carriage slams onto the tracks.' Then, immediately show Queen Victoria
stepping onto the train, skipping the date. Build each scene with tight,
visual action: advisors panicking, crowds gasping, Victoria gripping her
seat, the train lurching, her dress whipping, and end with the pop-culture
twist and the final awe-struck question.
reasoning: The script fails the **Opening Hook** audit due to a lack of a
clear, concrete visual subject before the timeline drop. This caps the
score at 60. While the **Tension & Escalation** and **Historical Surprise**
are strong, and the **Final Impact** is solid, the initial line is confusing
and abstract, which will lose viewers instantly. The TTS cues are energetic
and cinematic, but the script needs a sharper, more visual start to
maximize retention.
focus: Opening Hook
Then followed a two more iterations to reach scores of 92 and finally 96 with:
Steam explodes as a royal carriage slams onto the tracks.
[Hit the gas with cinematic, explosive hype—like a scene in a wild
blockbuster trailer]
Queen Victoria steps out, skirts swirling, eyes locked on the smoke-belching
train.
[Snap forward—vivid awe, zero delay, make the viewer feel like they’re in
the crowd]
Advisors freak out. One nearly faints. The crowd literally gasps as she
grabs her seat.
[Lean in with giddy disbelief—like you’re sharing secret backstage drama]
The train jerks forward. Victoria grips the rail, her dress whipping in the
blast of steam.
[Ramp up like a thrill ride—show movement, lift tension, put wind in her hair]
It’s like if Beyoncé livestreamed her first rocket ride—every headline
explodes.
[Land with an epic pattern-break—modern pop-culture hit, pure main character
energy]
Suddenly, everyone wants to ride the rails.
[Fast montage, energetic—like headlines flashing and crowds buzzing,
newspapers flying off stands, copycats lining up at the station]
Suddenly, trains are the symbol of cool. Who else would’ve dared to ride first?
[Wrap in awe, voice soft with wonder but eyes wide—leave them hungry for more]
G. Illustration
With the voice-over done, it’s time to illustrate those 7 scenes with the next pair of Beta-Gamma AI agents. We collect the prompts from Github and execute the Beta AI agent to create the illustrations, here is the first iteration for the first scene:
scene_role: Hook/Thumbnail
image_prompt: Queen Victoria, Windsor station, June 13, 1842, morning.
Foreground: A close-up of Queen Victoria's face, wide-eyed and defiant,
reflected in the glossy window of a royal carriage as scalding steam billows
around her. Mid-ground: The royal carriage slamming onto the train tracks,
wood splinters flying, metal wheels striking rails, an explosion of white-hot
steam and clouds. Background: A surging Victorian crowd, top hats and bonnets
tossed by the blast, ornate station arches visible through the haze, Union
Jack flag partly obscured by smoke. Low-angle, cinematic close-up, dramatic
depth of field. The queen’s gloved hand clenches the edge of the window, a
golden locket glinting at her neck. Motion: billowing steam, skirts
fluttering, a startled pigeon takes flight. Color: rich dark blues and
golds, vivid red cloak, white steam. Hidden detail: A startled mouse peeking
from the rails. 9:16 vertical. Do not generate images with extra limbs,
distorted anatomy, or unrealistic features. Characters must have 5 fingers
per hand, proper proportions, and no duplicated body parts. Use the Digital
Cinematic Concept Art visual style—modern digital rendering, cinematic
lighting, and stylized realism. Do not use brushstroke or canvas textures.
backup_image_prompt: A symbolic scene: A gleaming golden crown balanced on
the edge of a roaring, steam-blasting locomotive, with billowing smoke
forming the edges of a Victorian-era map. Victorian crowd silhouettes line
the platform. Low-angle, bold color contrast between royal blues and fiery
steam. Motion: smoke swirling, train wheels spinning. Hidden detail: a tiny
Union Jack woven into the steam. 9:16 vertical. Do not generate images with
extra limbs, distorted anatomy, or unrealistic features. Characters must have
5 fingers per hand, proper proportions, and no duplicated body parts. Use the
Digital Cinematic Concept Art visual style—modern digital rendering,
cinematic lighting, and stylized realism. Do not use brushstroke or canvas
textures.
video_prompt: Quick push-in from the station floor toward Queen Victoria’s
face inside the carriage as steam explodes around her, then whip-pan to the
slamming of the carriage wheels onto the tracks. Freeze on her wide-eyed look,
steam swirling. Upright 9:16.
overlay_subtitle: Queen Victoria rides a steam train
On this run, Gamma judge the score at 95, so there will be not 2nd iteration, but still suggested:
suggestion: To reach a perfect 100, rewrite the climax and/or aftermath to
embed the YouTube CTA more literally and era-appropriately—e.g., show a
Victorian newsboy waving a headline that reads 'Subscribe for more royal
rides' in period font, or have a telegram or newspaper in the aftermath scene
where the subscribe/thumbs-up icon appears as a woodcut or headline motif,
always upside-down to the viewer if being read by a character. This will
maximize both shareability and period immersion.
H. Build Metadata
Now that we have almost everything, we need to define the YouTube metadata (title, description, playlists, etc.), the background music, a Tweet message to post about the video, a question and an URL to add as comments on the YouTube Video and details to build a commemoration T-Shirt on Fourthwall (check out the-ai-illustrated-history-shop.fourthwall.com). The AI agent provided the following for this run:
title: Queen Victoria’s Wild Train Ride: When Royalty Shocked the World 🚂 #june13 #history
description: #History #OnThisDay #1842 Steam blasts, crowd gasps: Queen Victoria takes the ultimate leap into the unknown! On June 13, 1842, the young monarch shattered centuries of tradition by boarding a roaring steam train—the first British royal ever to do so. Her bold ride turned the world’s most feared machine into a symbol of progress, inspiring millions to embrace modern tech. Find out how one royal train trip changed travel forever and why her main character energy still echoes today. Want the full story? Check the link in the comments!
background_music: a_new_leaf_silent_partner.mp3
playlists: European History Shorts, Science & Discovery, Inspiring Stories
filename: queen-victorias-first-train-ride-june-13-1842.mp4
tweet: Imagine if Beyoncé took the Hyperloop live—on this day in 1842, Queen Victoria did just that with steam trains! She braved doubters, shocked crowds, and made railways the new hot thing. Hero move or wild gamble? #OnThisDay #history #June13 🚂
url: https://en.wikipedia.org/wiki/Queen_Victoria%27s_first_train_ride
question: Would you risk your reputation on a brand-new technology—knowing the world is watching? Main character energy or too risky?
merch_image_prompt: Create a bold, stylized 16:9 illustration showing Queen Victoria in regal attire gripping the railing of a roaring Victorian steam train, wind whipping her dress, clouds of dramatic steam billowing around. The train streaks powerfully across the tracks, a golden crown glinting at the front, as Victorian-era headlines swirl past in the air (“QUEEN RIDES THE RAILS!”). The scene uses deep royal blues, gold, and crisp white for maximum contrast on a dark t-shirt, and the queen radiates confidence and defiance—main character energy. Add a witty slogan in large, bold modern font: ‘RIDE IT LIKE A QUEEN’. Ensure ample padding: the image and the text have at least 10% border distance from the edge—no words or key design elements near the edge. No brands, logos, or YouTube elements
merch_image_subtitle: "June 13, 1842 – Victoria’s Epic Train Flex
merch_title: Ride It Like a Queen – Victoria’s Bold Train Leap
merch_description: Level up your style and your courage: ‘Ride It Like A Queen’ celebrates June 13, 1842, when Queen Victoria made royal history and turned steam trains from fearsome machines into icons of cool. Rock this ultra-soft, heavyweight tee featuring Victoria’s gutsy first train ride and a slogan that drops pure main character energy. Pairs perfectly with bold choices and everyday adventures. 100% ring-spun cotton, heavyweight, pre-shrunk, and always comf
I. Store Data
Now that we have everything, n8n store all the information in Google Drive, record some tracking information on BigQuery.
J. Assets creation
n8n now jump into a new sub-workflow that will iterate on each scene and create the voice-over audio file with GPT-4o-mini-TTS and the image with either Leonardo, GPT-4o or Google Imagen. We are creating the image for the merch T-Shirt and on some occasion the thumbnail for long format video. All files get stored on the same Google Drive folder.
H. Video Editor
This custom-build application in Python, download all the assets and start developing the animation of each video. Alternatively zooming in and then zooming out. I am using a saliency detection from OpenCV to identify point of interest in the picture, a detail in the picture that human eyes will have the tendency to catch-on. Zooming in into those details create a natural effect which give a good experience. Setting up the scene time to the voice-over audio file. Adding the background music track on loop. Adding title and subtitle per guidance from the AI agent. Finally, the video is same in the specific filename define by the AI agent — SEO is king and even filename matters. At the end, the video get uploaded to YouTube with a release date 1h in the future, and two comments are being added (one open-ended question, and one comment with more information on the story).
From candidate scrape to public Short: ~12 minutes wall‑clock, of which my human input is… zero.
6. What the Numbers Really Say
After 322 Shorts, to provide you the good, the bad and the ugly, I compared the top 25% (tier A) with less performing video (tier C, between 50 and 75 percentile). Comparing those two groups, we see the following:
- Scene‑Level Retention. Top videos keep 95 % of viewers through Scene 1, while Tier C drops to 44 % in those same ten seconds. Hook delivery, not just headline wording, decides survival.
- Average View Duration. Tier A posts average watch times >100 % (rewatches inflate the metric); Tier C stagnates at 40‑50 %.
- Traffic Sources. Winners ride the Shorts feed (95 % of first‑week views). When the algorithm shrugs, Tier C leans disproportionately on Search (one C video logged 44 % of its meager views via search).
- Demographics. Both tiers skew older male, but Tier A occasionally snags under‑45 viewers (up to 36 %), hinting that broader appeal is possible.
- First‑Week Trajectory. Tier A spikes on 1st day yet still earns trickle views through 3rd day; Tier C often flat‑lines in 24 h (e.g., 561 views 1st day → 0 in the next 6 days).
Bottom line: If Scene 1 hemorrhages viewers, the video is doomed; fix the opening or don’t bother shipping.
Most Views: https://www.youtube.com/shorts/t5VMI53HVDs
Most Likes: https://www.youtube.com/shorts/pryx-eJxxo4
7. Closing Thoughts
AI didn’t sideline me; it multiplied my output. Treating models as teammates-not crystal balls-let me publish a few hundred history lessons and iterate at breakneck speed. The data proves my most successful Shorts earn their keep in the first scene: hold attention fast, then ride the Shorts feed wave. Everything else-titles, tags, even thumbnails-is secondary.
If you’re eyeing your own AI‑driven side gig, remember: log every decision, iterate like a scientist, and keep a human hand on the wheel for final sense checks. History may repeat itself, but with the right prompts, it can also go viral.
Curious — check out www.youtube.com/@AIIllustratedHistory and my merch store the-ai-illustrated-history-shop.fourthwall.com