Tibo@tibo_maker1d ago

📝 how to build and run an army of AI UGC accounts in this article I'll walk you through all 9 steps of the setup, with the 12 prompts I use, ready to copy, in the order you'll need them. you've probably scrolled past a few of them this week. a girl talking through her skincare in the car, another woman showing what's in her gym bag, someone getting ready on a plane an hour before landing. they look like normal people filming normal parts of their day, which is exactly why you watched. a lot of them aren't real people. they're AI creators, and some brands now run several of these accounts at once, each one with its own face, its own life and its own audience. most people picture AI UGC as one fake face reading a script about a product, and that version is easy to spot and easy to skip. the version I'm talking about is different. it's a small group of creators who each have a believable life, and your product shows up for a few seconds inside that life, the same way a real person mentions something they use. this article is the full setup for running that with an AI agent: how to design the creators, starting from who actually buys your product how to keep their faces exactly the same across hundreds of videos how to find what's already working in each creator's niche how to turn those winners into original videos in each creator's voice how to keep every account posting on a schedule, with a review window before anything goes live how to read the results every week and tune the whole thing every prompt is in here, ready to copy. there are 12 of them, and they're numbered in the order you'll use them. here's the whole thing, in the order you'd set it up. why UGC works, and why it's so hard to run UGC works because it doesn't look like an ad. a normal person talking about something they use gets watched, and a brand video gets skipped. nobody really argues with that part anymore. the hard part is running it, and it breaks in the same four places. 1. it's an operations job. you're finding creators, briefing them, chasing them for drafts, sending notes, approving, paying, and then doing it all again with the next group, because creators rotate. most of the work isn't creative at all. it's managing people, and it scales with headcount. 2. you pay per video. every new hook is another invoice. every variation is another invoice. a reshoot because the lighting was off is another invoice. that would be fine if you only needed a few videos, but testing is the whole game on short-form. you find the hook that works by trying a lot of hooks that don't, and per-video pricing makes every one of those tries expensive, so testing is usually the first thing that gets cut. 3. it's slow. a creator needs a brief, a few days to film, and a round of edits. by the time the video comes back, the trend it was riding has often moved on, and you're posting last week's format into this week's feed. 4. you don't own it. the face, the account and the audience all belong to the creator. usage rights expire. when the creator stops working with you, everything they built while promoting your product leaves with them, and you start again from zero. put those four together and it turns into a loop: find a creator, brief them, wait, do edits, pay, they stop posting, start again. what changed three things are true now that weren't true a year ago. AI faces stay the same now. until recently, an AI person looked slightly different in every clip. the jaw shifted, the hair changed, the freckles moved around, and no audience could form around someone who kept turning into someone else. with saved characters you get the same face, hair and skin across hundreds of videos, which means people can actually recognise and follow a creator. you can search what's working instead of guessing. Revid's viral search lets you look up any topic and get the top-performing short videos on it, with transcripts, hooks and engagement data. that used to be a strategist's job: scrolling, saving, transcribing and writing up why something worked. now it's one search. agents can run the whole loop. an agent like Claude, Hermes, OpenClaw or Squad can call these tools directly. it can search, pick a reference, write the script, render the video and schedule it, then report back to you with what it did. put those three together and one person can now run what used to need a small team. the model: creators you own the idea behind this whole article is simple. instead of renting creators, you build a few of your own. each one is a believable person with a face that never changes, a normal life, and a set of topics they'd actually post about. your product shows up inside that life, and it's never the subject of every video. each creator gets their own accounts and their own daily pipeline. your agent runs all of them, and you review what's going out before it goes live. one creator is a channel. three or four of them, each speaking to a different kind of buyer, is the army. this fixes all four problems from before. there's no one to manage, because the agent does the work. there's no creator invoice per video, just render credits, so testing a new hook means one more render instead of one more brief. it's fast, because research to scheduled post happens in one run. and you own everything: the face, the accounts and the audience. who this works for (and who it doesn't) before you build anything, check the fit. this works best when: your product is used in everyday life (consumer apps, skincare, food, fitness, travel, productivity tools) the "aha" can be shown in a few seconds on a phone your buyers already watch short videos to decide what to try it works badly when: the purchase depends on deep trust in a specific person (medical, legal, financial advice) your category is regulated, or the platforms restrict claims in it the product only makes sense after a long explanation if you're in the second list, you can still use the research and scripting parts of this system for your own content. just skip the AI faces. the system the whole thing is six jobs, and each one can be done by more than one tool. the brain is the only part that touches everything else. it reads each creator's file, calls the research, production and distribution tools in order, and writes back what it did. if you swap one tool for another later, the brain and the prompts stay the same. the rest of this article is the setup, in nine steps. step 1: connect Revid to your agent what it does: gives your agent the ability to actually call the tools, instead of just talking about them. why it matters: an agent that can only write you a plan still leaves you doing the work. an agent that can search, render and schedule does the work and shows you the result. how to do it: Revid has an MCP server, so any agent that supports MCP can use it directly. in Claude Code it's one line: claude mcp add --transport http revid https://www.revid.ai/api/mcp then run /mcp in Claude Code, pick revid and choose Authenticate. other MCP clients use the same server URL with your API key. once it's connected, your agent can search viral videos, manage characters, render videos and schedule posts. then give your agent its brain file. this is the standing brief it reads before every single run. it tells the agent what the product is, who the creators are, what the daily job looks like, and the rules it can never break. you write it once and only touch it when something changes. keep it next to one file per creator, so the agent always knows where to look: ugc/ brain.md creators/ priya.md jess.md liam.md layla.md prompt 1: the agent brain file you run a small team of AI UGC creators for [PRODUCT]. product: [one line on what it does and who it's for] creators: [name and one line for each]. each creator has a file with their persona, topics, past posts and results. your job, every day: for each creator, pick one topic, find what's working on it with Revid viral search, write an original script in their voice, render it with their saved character, and schedule it on their own accounts at least 24 hours out. rules: - every account says it's an AI creator in the bio - never write fake reviews, testimonials or before/after claims - no health, income or medical claims - the product appears inside the creator's life, never as the whole video - never publish directly. schedule only, so I can review - log everything you do in the creator's file the two rules that matter most are "schedule only" and "log everything". the first one is what gives you a review window. the second one is what lets the agent avoid repeating itself and learn from what it posted. step 2: design your creators from your buyers what it does: turns what you know about your customers into a small set of creators, one per type of buyer. why it matters: the most common mistake is starting with a face. a nice-looking creator who speaks to nobody in particular gets views and no customers. if you start with who buys, every creator has a reason to exist. how to do it: split your buyers into a few segments that use the product for different reasons, then design one creator per segment. a wellness person, a frequent traveller, an outdoorsy guy and a city girl can all use the same serum, and each one reaches people the others never would. for the examples in this article we'll use four creators for a vitamin C serum called LumiGlow: Priya is a wellness creator in India. she posts about slow mornings and simple routines, and gives honest updates on her skin. Jess is a travel creator in the USA. she's always packing and always on a plane, so her skincare happens on the move. Liam is an outdoors creator in Australia. his life is surf, the ute and the van, and he's the guy who owns two products total. Layla is a lifestyle creator in Dubai. her life is the city, pilates and weekend trips to the desert, and her skincare is part of getting ready. notice that none of them is "a skincare creator". the serum fits into each of their lives, but it isn't what their lives are about. prompt 2: from buyer segments to creators here's my product: [description]. here's what I know about who buys it: [customer notes, reviews, survey answers]. split my buyers into 3 or 4 segments that use the product for different reasons. for each segment, design one creator who would naturally reach those people: - name, age, city, what they do - what a normal week looks like for them - 3 things they'd post about that have nothing to do with my product - the one moment in their life where my product shows up naturally keep them believable and ordinary. no influencer stereotypes. what you get: 3 or 4 creator outlines, each tied to a buyer segment, each with a normal week, three non-product topics and one natural product moment. pick the ones that feel most real and move on. then write each creator's bible. this is the document that makes every script sound like the same person, whether it's their first video or their two-hundredth. prompt 3: the creator bible write the full bible for [creator] so every script sounds like the same person. include: how they talk (sentence length, slang, words they'd never use), what they care about, their recurring places (home, work, the spots they go every week), habits and running jokes, what they wear, how they film (selfie, handheld, phone propped up), and what they'd never say or do on camera. end with 5 example opening lines in their voice. the bible goes at the top of the creator's file. here's the shape of a full creator file once everything in this article is set up, using Jess: # jess persona: travel creator in the USA, always packing, always on a plane, skincare happens on the move. bible: [output of prompt 3] character: [her saved Revid character] accounts: [her own TikTok, Instagram, YouTube] topics: [output of prompt 6, product beats marked] posting time: [her usual time] log: [date, topic, reference, script, video link, post time] results: [weekly numbers from the review] step 3: give them a face that never changes what it does: builds each creator's face once and locks it, so every video after that uses the same person. why it matters: consistency is everything here. if your creator looks a little different in every video, nobody recognises them, and recognition is the reason people follow an account in the first place. how to do it: build the face once, from a detailed description. the details in the prompt below are there for a reason. one distinctive detail like freckles or a small scar is what makes a face recognisable at thumbnail size. natural skin texture with visible pores and no retouching is what keeps it from looking like an AI render. arm's-length framing that's slightly off is what makes it look like a phone photo. prompt 4: the base portrait a realistic smartphone photo of [creator]: [age], [heritage], [hair: colour, length, texture], [face: shape, eyes, one distinctive detail like freckles or a small scar], natural skin texture with visible pores, no retouching. [setting], [natural light source], shot on a phone at arm's length, slightly imperfect framing. [outfit]. relaxed, candid expression. generate a few, pick the one that feels like the person in the bible, then turn it into a character sheet so the model has every angle to work from. prompt 5: the character sheet using this exact person as the reference, create a character sheet: front, three-quarter left, three-quarter right, profile, and one full-body shot. three expressions: neutral, laughing, and talking mid-sentence. keep the face, hair, skin texture and build exactly the same in every frame. plain light background. check the sheet frame by frame before you save it. if the face, hair, skin texture or build drifts in any frame, regenerate that frame. whatever gets saved here is what every future video inherits. then save it as a character in Revid, so every render after this uses the same person. if you're using a different image model, the character sheet itself is your reference, and you attach it to every generation. step 4: give each creator their topics what it does: gives each creator a list of things they'd genuinely post about, and marks the few where your product can appear. why it matters: a creator who only talks about your product is an ad with a face. people follow creators for their life, so most of what they post should have nothing to do with you. the non-product videos are what build the audience. the product videos are what that audience is for. how to do it: start each creator with a short list, and mark where the product fits: then have the agent expand it into a full month, with a search query for each topic. the search query matters as much as the topic, because it's what the agent will use in step 5 to find what's already working. prompt 6: from creator to topics and search queries here is [creator]'s bible: [paste]. list 20 topics they would genuinely post about in a normal month. at least 15 should have nothing to do with [product category]. for each topic, write the search query you'd use to find viral short videos on it: plain words a viewer would type, 3 to 7 words. mark the 5 topics where [product] could appear naturally for one beat. what you get: 20 topics per creator, each with a 3 to 7 word search query, and 5 of them marked as product beats. that's a month of posting for one creator. it goes into their file under topics. step 5: find what's already working what it does: for every topic, finds the short videos that already performed and breaks down why. why it matters: don't start from a blank page. for every topic, someone has already made a video that worked, and the structure of that video is the most useful thing you can learn from. you're not guessing what the hook should be. you're looking at hooks that already worked on a real audience. how to do it: your agent searches each creator's topics and looks at the top results: the hook, the structure, the pacing, the camera style, and why people watched to the end. with Revid connected, that's the search_viral_videos tool. it returns the top videos for a query across TikTok, Instagram, YouTube and more, with transcripts, engagement and style labels like selfie, handheld or tutorial. you can do the same thing by hand in TikTok search, it just takes a lot longer. two settings make a big difference: sort by most popular, with viral-only switched on. this cuts out everything that's just recent and leaves what actually spread. filter for person-focused, selfie-style videos. that's what UGC looks like, and it's what your creators will be making. a winning video with a voiceover over stock footage teaches you very little about how your AI creator should film its morning. open any result and you get the detail panel: plays, the full transcript, and a breakdown of why it works. then hand the results to your agent and have it break down the winners properly. prompt 7: breaking down the winners here are the top videos for "[query]", with transcripts and engagement: [paste results]. for each one, break down: the hook (first 2 seconds, words and visual), the structure beat by beat, roughly how long each beat runs, the camera style, and why someone would watch to the end. then pick the one that fits [creator] best and explain why in two sentences. prefer one continuous, selfie-style video over edited lists. what you get: a beat-by-beat breakdown of every top video, and one pick for the creator with a two-sentence reason. the last line of the prompt matters. edited lists are hard to make feel like a real person filmed them, while one continuous selfie take is exactly what a real person would post. step 6: turn the winner into their video what it does: rewrites the winning video as an original video by your creator. why it matters: this is the line between learning from a format and copying someone's video. you want the first one. the structure of a video is a pattern anyone can use. the words, the story and the person are someone else's. how to do it: the rule for remixing is keep the structure, change everything else. the result is clearly an original video by your creator, built on a structure that's already proven to hold attention. the product goes into one beat, the way a real person mentions something they use. no claims, no discount codes, no "link in bio". if you read the script out loud and it sounds like the creator is being paid, it's wrong. prompt 8: the remix script rewrite this video as [creator]. original transcript: [paste] keep: the hook type, the structure, the pacing, the length. change: every word, the setting details, and the personal story, so it's clearly theirs. place [product] in exactly one beat, the way a real person mentions something they use. no claims, no discount talk, no "link in bio". write it in their voice from the bible. spoken, not written. under [length] seconds. what you get: a spoken script in the creator's voice, on the same structure as the winner, with the product in exactly one beat. you'll see a full example in the Jess run below. step 7: put it on a schedule what it does: runs steps 5 and 6, plus rendering and scheduling, for every creator, every day, without you starting it. why it matters: this is where it becomes a system instead of a pile of prompts. everything up to here, you could do by hand once. the daily run is what makes it an army. how to do it: once a day, your agent runs the whole loop for every creator. it reads their file, picks a topic, searches, picks a reference, writes the script, renders with their character, and schedules it on their accounts. then it sends you one message with everything it queued. this assumes each creator's accounts are already connected, so the agent has somewhere to schedule to. setting those up is covered in step 8. every step maps to a Revid tool, so nothing here is hand-waving: you can run the schedule from whichever agent you use: Claude's scheduled tasks, a cron job for OpenClaw or Hermes, or a scheduled task in Squad, which is what I use. this is the most important prompt in this article. it's the one that runs every day. prompt 9: the daily run it's the daily run. for each creator: 1. read their file: persona, topics, the last 14 posts, what performed. 2. pick today's topic. don't repeat a topic used in the last 10 days. 3. call search_viral_videos with that topic's query. platforms tiktok and instagram, person-focused, viral only, sorted by popular. 4. pick one reference you haven't used before. note why. 5. write the script with the remix rules. 6. call calculate_credits, then render_video with the creator's characterIds. poll get_project_status until there's a videoUrl. 7. call schedule_publish to that creator's own accounts, at least 24 hours from now, at their usual posting time. 8. write the reference, script, video link and post time to their file. when you're done, send me one message: each creator, the topic, the reference you used, the scheduled time, and the video link. context on this prompt: "the last 14 posts" gives the agent enough history to avoid sounding repetitive, without loading the whole archive every day. "don't repeat a topic used in the last 10 days" keeps each creator's feed varied. "one reference you haven't used before" stops the agent from remixing the same winner over and over. "calculate_credits" before "render_video" means you know the cost before anything gets made. "at least 24 hours from now" is your review window. more on that in step 8. step 8 of the prompt is what makes tomorrow's run smarter than today's, because everything it did is now in the file it reads first. the army in action here's one run for Jess, start to finish. the ask. I asked for what's going viral on "skincare routine on a long flight" and a pick for her next video. the search. the agent searched TikTok and Instagram, person-focused, viral only, sorted by popular. the top five results had 6.6M, 5.6M, 3.1M, 2.2M and 2.1M plays. the pick. it picked a get-ready-with-me video filmed in one continuous selfie take on a plane. it skipped the two most-viewed results: the 6.6M one is a 7-second joke with no routine in it, and the 5.6M one is a list of travel hacks, which is much harder to make feel like Jess. that's the "prefer one continuous, selfie-style video" rule from prompt 7 doing its job. the script. then it wrote her script on the same structure, with the serum in exactly one step: 13 hours in the air and I refuse to land looking like it. here's what I do in the last 30 minutes before we touch down. face wipe first, nothing fancy. then two drops of LumiGlow, because plane air dries me out and it's the one step I never skip. moisturiser on top, lip balm, a little concealer only where I need it. hair up, hoodie stays on. landing in 20. see you on the ground. look at where the product sits. it's step two of six, it gets one reason that's about her life (plane air), and then the video moves on. there's no claim about what it does to your skin, no code, no link. the video is about landing on a long flight without looking like it, and the serum is just part of how Jess does that. the same thing happens for Priya, Liam and Layla every day, each with their own topics. here's what a week of Liam's posts looks like once it's queued: step 8: accounts and the review window what it does: gives every creator their own home on each platform, and gives you a daily window to catch anything before it goes live. why it matters: the whole point of this system is that people follow a person, and that only works if each person has their own accounts. scheduling 24 hours out is what keeps you in control without approving every video by hand. how to do it: each creator gets their own TikTok, Instagram and YouTube. never mix two creators on one account. Priya's followers followed Priya, and a Liam video in her feed breaks that. say they're AI in the bio. it costs you less than you'd think, and it protects every account you build. prompt 10: profile setup write a profile for [creator] for TikTok, Instagram and YouTube: display name, 5 handle ideas, and a bio under 80 characters in their voice. the bio must say they're an AI creator in a natural way, for example: "AI creator. carry-ons, long flights, too many airports." what you get: a display name, 5 handle ideas and a short bio per creator, with the AI disclosure written in their own voice instead of a legal-sounding line. the review window. the review window is the 24 hours between the agent scheduling a video and it going live. you don't approve every video one by one. you look at the queue once a day and remove anything that's off: a face that drifted, a script that leans too hard on the product, a topic that's too close to last week's. that's how you stay in control without becoming the bottleneck again, which was the whole problem with running UGC the old way. step 9: measure and improve what it does: once a week, reviews every creator's posts, decides what to do more of and what to drop, and updates their files. why it matters: views alone will mislead you. what matters is whether people watch to the end, save it, and tap through to the profile, because that's what turns a viewer into a follower and a follower into a customer. how to do it: once a week, your agent reviews each creator: what worked, what to drop, and what to test next. give it the numbers from your analytics, or connect it to them if your setup allows. Revid's get_publish_status confirms a post went live, but the performance numbers come from each platform's analytics. prompt 11: the weekly review it's the weekly review. here are the stats for each creator's posts from the last 7 days: [paste or pull]. rank posts by watch-through and saves, not just views. for each creator, tell me which topics to post more of, which to drop, and 5 new topics or search queries to test. update each creator's topic list and file. if a creator has had 3 flat weeks in a row, flag them so we can rethink the persona instead of posting more of the same. what you get: a ranked list of the week's posts per creator, the topics to double down on, the ones to drop, 5 new topics or queries to test, and updated creator files. the next daily run reads those files, so the review changes what gets made the very next day. the "3 flat weeks" line is there for a reason. when a creator isn't working, the instinct is to post more. usually the persona is the problem, and more posts of the wrong person just gives you more of the same result. and when one video underperforms, don't rewrite everything. find the one thing that broke. prompt 12: when a video flops this video underperformed: [link, stats, script, reference used]. compare it to the reference and to this creator's best post. was it the hook, the topic, the length, the moment the product appears, or the visuals? give me one change to test next time, not five. what you get: one diagnosis and one change to test. one change at a time is the only way to know what actually fixed it. what to watch out for none of this runs perfectly on its own. these are the things that go wrong, and what to do about each one. product-first scripts. if the product is the topic of every video, the account turns into an ad and people stop watching. keep it to one beat, and keep most videos about the creator's life. the 15 out of 20 rule from step 4 is what protects this. too many topics per creator. a creator who posts about everything reads as nobody in particular. stay inside their world. Jess talks about flights and packing, and she doesn't suddenly start reviewing protein powder. views without buyers. a video can go viral and sell nothing. watch profile visits and clicks, not just plays. the full setup connect Revid to your agent like Squad and give it the brain file split your buyers into segments and design one creator per segment build each face once and save it as a character give each creator their own topics and search queries find what's already working on each topic remix the winner into their video, with the product in one beat put the daily run on a schedule one creator, their own accounts, AI in the bio, and a 24-hour review window a weekly review that tunes topics and flags weak creators start with one creator. get their videos right, then add the next. if this was useful: - repost it for anyone still paying for every single video - follow me for more on running marketing with agents tools used in this setup: Revid.ai and Squad.so https://x.com/i/article/2113633101888847874