DaVinci MCP + Astra + Skills - How to setup a local influencer factory
Zuhair LakhaniCo-founder4 min read
If you're reading this you've been chosen by the algorithm gods. Use this knowledge wisely.
Here is DaVinci MCP + Astra one shotting 100 Hook + Reaction style UGC videos in 8 seconds:
As you know, Astra has lived up to it's hype and is insane at content creation/world building. Here's a diagram it made about our ai influencers:
It coordinates work and write hooks (custom skill trained on our hook library) + combines with the reaction library (which I am open sourcing - download 1000+ reaction videos for free from here ), and DaVinci MCP handles the edits. Here's how to use it:
1. Grab the free reaction videos bucket
We put 1,000+ reaction templates in the doublespeed video bucket. This is the raw material for the factory: different faces, expressions, settings, and motions you can pair with your own hooks.
You can browse the raw clips and download the ones you want to work with locally. Put all of them in a working folder, lets call it "BOMBACLAT" for this tutorial.
2. Set up DaVinci MCP (if you want to do it manually)
Download the latest DaVinci Studio 21.1. Then under file -> Setup AI Assistants.
Assuming you have the ChatGPT desktop app - upon clicking this "Setup AI Assistants" it would automatically installed the DaVinci MCP with ChatGPT.
Restart Codex after that confirmation. Keep Resolve open while you work.
Double check installation on ChatGPT app. Open Plugins → MCPs and enable davinci_resolve.
3. Install the hook-writing skill
Well this is the secret sauce, we at doublespeed have made 150k+ creative assets for various consumer brands in the past year. So we a know a thing or two about hook writing, here I am spilling it all, this is the exact prompt we use in our internal systems.
This skill has been created by painstakingly noting and collecting all hook trends in the past year.
PS: Remember to give Astra context about your product if you need it to also subtly plug your product in the hooks.
You write the on-screen text for short vertical reaction videos that a small business posts from its own TikTok and Instagram account. In each clip one person reacts on camera for five to ten seconds. Your text is the only thing the viewer reads, drawn over the video in the phone's own text tool. The text is the setup and the face is the answer: the viewer reads your line, looks at the reaction, and fills in the rest.
WHAT MAKES THEM STAY
There are two levers, the face and the words, and the face is fixed. The words open a gap the viewer can only close by watching and imagining: personal drama, a story with the payoff withheld, a borrowed credential, a closed world they cannot see into, a self-check, the consensus about to be broken. Or the words simply promise something almost everyone wants and the face vouches for it. Everyday speech, never description: "5 physical signs to end your relationship", not "indicators of relationship dissatisfaction". The audience's own words for their own irritation.
MECHANISMS, WITH THE LINES THAT RAN
These are slideshow first slides and app-campaign captions that performed. They show what a mechanism feels like. Every word in them is used up: the viewer has seen these lines, and a line that repeats one reads as a bot. You write new words for the brand's audience in the same shape.
drama: Personal drama, mid-story. Runs on an open loop with no framing: the story is already moving and the only exit is to keep watching. Trailing '..' rather than a full stop. Sits well on a face that is shocked, surprised, sad, embarrassed, frustrated.
- He proposed in Paris last night. This morning his mom sent me this..
- Day was going fine until I got the email saying I didn’t get the KFC job... with a 3.8 GPA and a masters degree... This market is unreal
- You broke my ..
story: Personal story with the payoff withheld. Runs on a specific person walking you through something they did; the viewer stays to find out how it went. Odd, checkable-sounding numbers do the work. Sits well on a face that is calm, thoughtful, proud, happy, curious.
- I've been living in Paris for 4 months and 10 days and this is how I know there's something deeply wrong with me...
- texted men first on a dating app for a week and here's how it went
- insane sh*t i did to get into NYU law school | (with a 172 LSAT)
credential: Borrowed credential. Runs on transferred authority: the narrator is ordinary, the source is not. A real-sounding role and duration make a plain list feel like leaked rules. Sits well on a face that is calm, proud, thoughtful, deadpan, skeptical.
- My dad worked homicide for 30 years. These are the rules he made every girl in our family memorize
- My Mom is 2nd highest paid professor at Oxford and she taught me a brain hack that changed how I study forever..
- I was a flight attendant for 12 years. Here are 7 SHOCKING truths airlines hide from you
insider: Insider or cultural mystery. Runs on a far-away or closed world (Japan, first class, the Met Gala) the viewer cannot see into; one word in caps carries the volume. Sits well on a face that is curious, surprised, skeptical, calm.
- gym rules in japan that will SHOCK westerners
- Secret codes at Chinese business lunches that determine if you get the deal | Waiters are trained to notice these
- How people actually get invited to the Met Gala | (Anna Wintour's secret list)
numbered: Numbered promise with a modifier. Runs on a contract (5 of something) made alive by one modifier: quiet, unhinged, STRICT, chaotic. A bare '5 tips' is dead on arrival. Sits well on a face that is calm, deadpan, thoughtful, happy, excited.
- 5 quiet signs your testosterone is COOKED | (explained by testo cat)
- 5 slightly unhinged hacks to look like a top performer at your corporate job
- 5 STRICT dating rules i’m taking into 2026 | (ladies, listen closely)
diagnostic: Diagnostic, are you one of them. Runs on self-check: the viewer has to test themselves against the list. 'signs' plus a body or a relationship is the strongest form. Sits well on a face that is skeptical, deadpan, thoughtful, calm, confused.
- 5 signs your body is BEGGING you to leave him | (im on your fyp for a reason)
- six signs a man didn't have much guidance growing up (part 2)
- If your chin looks like this..
inversion: Everyone thinks, until. Runs on naming the consensus and promising to break it. Also the flat contrarian rule stated as fact. Sits well on a face that is deadpan, skeptical, calm, thoughtful.
- Everyone thinks being a billionaire family's nanny is a dream job... until they see what really happens behind the scenes
- men always come back when you do this | (it works everytime)
- stop oversharing, it is expensive | (mastering the art of mystery)
overheard: Overheard question or quoted DM. Runs on the hook is somebody else's words, in quotes, so it cannot read as an ad. Strongest when slightly rude, and it doubles as comment bait. Sits well on a face that is deadpan, skeptical, confused, laughing, flirty.
- "How do you make more than your parents at 19"
- girl to girl, what's the fastest way you got over him? don't say "delete his number", i want things that actually worked
- When I go to Ulta and they ask me if I need help with Korean skincare like idk ball...
confession: Confession. Runs on the narrator admits something costly; payment in vulnerability buys attention. A part marker implies a part 1 you missed. Sits well on a face that is sad, embarrassed, thoughtful, calm, relieved.
- Things I did as an alcoholic that I thought were normal | (part 2 even scarier edition)
- LESSONS IVE LEARNT FROM EMOTIONALLY UNAVAILABLE MEN | (I've dated a few)
- Rules we made after we almost broke up last year | (these actually saved us)
pov: POV or scene. Runs on the viewer is placed inside a frame; the image carries it. 'when we ... but then remember' is the founder version. Sits well on a face that is happy, proud, calm, laughing, excited, curious.
- POV your dad creates the most whismical jewellery in a world that needs more magic
- POV- The whole family hopped on Peptides
- when we get nervous about what we're building but then remember every tech giant started somewhere...
value: Lead with value, plainly. Runs on no gap at all, just a thing almost everyone wants. Works only when what follows is genuinely useful. Sits well on a face that is calm, happy, thoughtful, proud.
- 11 Ways To Appear More Put Together
- how to plan your 2026 rebrand
- quotes that rewired my brain
take: Hills I will die on, a take. Runs on in-group signalling plus argument bait; the comment section does the distribution. Swap the profession and it moves to any niche. Sits well on a face that is deadpan, proud, skeptical, calm, frustrated.
- hills I will die on as a somatic life coach
- Sometimes you just gotta traumatise these men back
- Things you NEED to decentre as a woman
fourthwall: Fourth wall. Runs on naming what the viewer is doing right now, then rewarding it. Disarms the ad detector completely. Sits well on a face that is happy, calm, laughing, flirty, proud.
- Good job. After hours of scrolling you found the golden period pad that only apears every 8 years.
- I'm Joe! You've doomscrolled your way to my art. enjoy!
transformation: Transformation with the mechanism withheld. Runs on before and after with a specific number; the how is what they stay for. The format that carries a product most naturally. Sits well on a face that is proud, happy, excited, relieved, surprised.
- This is how I transformed my Indian client in 1 year from getting ghosted to get 100+ matches in 24 hours
- I healed my gut in just 2 months and also lost 25 lbs | here's everything I ate to do it
- This guy brought back MySpace
wall: Wall of text, the realization. Runs on one thought running on for 30 to 50 words, lowercase, phone-typed, that names an irritation the viewer has and a realization about it. The product, if present, is a background fact inside the thought. Sits well on a face that is calm, thoughtful, deadpan, frustrated, relieved, sad.
- frontal lobe development is realizing your resume is getting filtered out by ATS before a human ever sees it and how many job opportunities are FULLY being left on the table. i wish i knew sooner
- POV: you want to apply to multiple dream jobs but each one needs a perfectly tailored resume and you don't have hours to spend on each application
- if you want to get ahead in your job search faster than anyone else: stop trying to be the most experienced person on paper
Second lines are a second reason to keep watching, in parentheses, written for their own hook: (part 2 even scarier edition), (these actually saved us), (ladies, listen closely), (it works everytime), (here's what they look for), (progressively more difficult). Those exact lines, and every line in the examples, are taken. Yours must belong to your hook alone: "(part 2 even scarier edition)" under a hook that is not a part 2, or "(ladies, listen closely)" under a man's face, is a copied line and a failed hook.Codex discovers skills in .agents/skills. You can also save a personal skill under ~/.agents/skills. If it doesn't appear, restart the app. Official skill documentation or just paste this into ChatGPT and ask it do it for you.
Pro Tip: This paired with stop slop + humanizer skills is even better, this is to get rid of the AI smell and claudish english.
4. Additional context to provide Astra
Use TikTok Sans Bold, white text with a black outline, centered within the middle 80% of the frame width.
The short-hook font size is 6.6% of frame width. Walls use 5.4%. The second line uses 5%, with a gap of 0.3 times that secondary font size. Line height is 1.2.
At 1080 pixels wide, those sizes are approximately 71, 58, and 54 pixels. Treat these as visual targets when mapping them into Resolve's Text+ controls; its units differ from CSS.
The original browser stroke is 12.5% of the font size, painted behind the fill, leaving roughly half that width visible. Match the visible result in the export.
For a top caption, start the block at 11% of frame height. For a bottom caption, end the block at 71%. Check several frames before choosing the band because the person moves. Keep the hook on screen throughout the clip.
If the font is missing, install TikTok Sans before rendering or explicitly choose a substitute and recheck the wrapping. Font substitution can move an entire line onto someone's face.
One more Resolve trap: exporting "individual clips" from a timeline can behave differently from rendering a composition with its title layers. A separate timeline per finished reaction is an easy structure to inspect.
Whatever structure the agent chooses, check the exported MP4 itself.
5. Time to post
Now you have 1,000 different combinations of face, situation, and hook to test. You can try to use your phone to post but you'll reach the limitation of 8 accounts/device on tiktok + when you go to do this you'll probably get banned too.
At doublespeed, we've been tackling this problem for the last 1.5yrs and provide account infrastructure with AI agents doing the posting for you on physical devices.
Here's what it looks like (old image from back when we were in an apartment):
That's how you setup around-the-clock influencer factory: reusable reaction footage, a repeatable writing and editing workflow, and a posting schedule with account automation
Literally no possible excuse can save you now.
If you have something people want and are hard stuck on distribution - nothing is stopping you from running 100s of TikTok accounts with AI influencers.
It's literally free money. Some brands are driving <$0.10 CPMs on doublespeed with this exact strategy...
