openai shipped its most capable model on september 3. i wasn't going to write this.
we've generated 2.6 billion organic views across 62 campaigns, managed 590 creators, and produced 22,504 tracked videos for mobile apps. zero ad spend.
these 15 prompts are how i'm using ChatGPT Astra to run those campaigns at a level that was physically impossible 2 weeks ago. every single one is copy-paste ready. i'm giving them all away because by the time you read this i'll already be running them on the next campaign.
save this. you will need it.
before you run a single prompt: load your campaign context
this is the step everyone skips and it's the step that makes everything else work. without it you're asking a stranger for advice about a business they've never seen.
paste this into ChatGPT Astra before anything else:
1"here is everything you need to know about my UGC program before we start any analysis. reference this every time i ask you to run an audit, generate a brief, or analyze performance. never ask me for this information again.23BUSINESS BASICS:4app name: [your app name]5app store URL: [iOS and/or Android link]6category: [health, education, finance, social, etc.]7monthly downloads: [X]8MRR: [if comfortable sharing]9one-sentence value prop: [what the app does in 15 words or less]1011CAMPAIGN HISTORY:12total creators managed: [X]13total videos produced: [X]14total organic views: [X]15platforms active on: [TikTok, Instagram Reels, YouTube Shorts]16average views per post: [X]17best single video performance: [X views, link if available]18biggest campaign win: [one sentence]19biggest campaign failure: [one sentence - be honest]2021CURRENT CAMPAIGN:22campaign name: [X]23number of active creators: [X]24videos per creator per week: [X]25brief template: [paste your current brief or describe it]26compensation model: [flat fee per video / monthly retainer / bonus structure]27target hook rate (1s): [X% or 'don't track yet']28target view-through rate (3s): [X% or 'don't track yet']2930CONTENT STRATEGY:31primary content formats: [testimonial, tutorial, day-in-my-life, reaction, before/after, etc.]32top 3 performing hooks from last 30 days: [paste exact hook text]33top 3 underperforming hooks: [paste exact hook text]34competitors running UGC: [app 1, app 2, app 3]3536HOW I WANT YOU TO WORK:37always prioritize actionable output over analysis.38when you give me a recommendation, tell me impact level (high/medium/low) and time to implement.39output data in table format when comparing anything.40never invent a number. if you can't verify it, say so.41when analyzing hooks, cluster by structure not by topic."
once this is loaded, every prompt that follows gets 10x sharper. Astra stops answering a generic marketing question and starts answering yours.
PART 1: VIDEO ANALYSIS (prompts 1-3)
everyone thinks Astra can't watch videos. they're wrong. it just can't watch them the way you think.
Astra takes text and images as input. it does not accept raw .mp4 files. but here is the thing most people miss: you don't need it to. you need a pipeline that watches the video FOR it and hands it a structured breakdown.
this is the system we actually run, and it's the technique that changed everything for us.
the pipeline: how we make AI "watch" 200+ videos per week
we built an automated video intelligence pipeline. here is how it works end to end.
- step 1: agent watches the video
we use an AI coding agent (openai codex, claude code, or any agent with tool access) connected to two command-line tools:
- yt-dlp: downloads any TikTok, Instagram Reel, or YouTube Short from a URL
- ffmpeg: extracts frames at adaptive intervals (more frames during the hook, fewer during the middle)
the agent runs this automatically. you give it a video URL. it downloads the video, extracts 8-12 key frames (weighted toward the first 3 seconds where hooks live), pulls the transcript from captions or runs whisper for speech-to-text, and reads any on-screen text from the frames.
the output is a structured report per video (btw this is a video from one of our campaigns):
1VIDEO: @juliislearning2 - TikTok3URL: https://tiktok.com/@juliislearning/video/76502986552362467924DURATION: 55 seconds5FRAMES EXTRACTED: 126FRAME 0:00 - creator sitting at desk, hands on notebook, text overlay7top-left: "3 things i wish i knew before finals"8FRAME 0:01 - eye contact with camera, slight head tilt, text still visible9FRAME 0:03 - cuts to phone screen showing app, finger scrolling10FRAME 0:05 - back to face, talking, hands gesturing11FRAME 0:10 - screen recording of app in use, flashcard generation12FRAME 0:15 - split screen: creator left, app right13FRAME 0:25 - close-up of app results14FRAME 0:35 - creator reaction face15FRAME 0:45 - back to talking head, summarizing16FRAME 0:52 - CTA frame, text overlay: "link in bio"17TRANSCRIPT: "ok so if you're a student and you haven't tried this18yet i genuinely don't know what you're doing. i found this app19called knowt like two weeks ago and it literally makes flashcards20from your notes automatically. you just upload a pdf and..."21ON-SCREEN TEXT: "3 things i wish i knew before finals"22CAPTION: "this changed my entire study routine honestly #studytok23#finals24 #studywithme25"
2. step 2: batch process all campaign videos
the agent doesn't stop at one video. we point it at an entire campaign. it processes every video URL from our sideshift export and generates a structured report for each one. for a 50-video batch, this takes about 20 minutes running in the background. no human involvement.
the output is one file with every video described in the same structured format: frames, transcript, on-screen text, visual elements, pacing, scene changes.
3. step 3: feed everything into astra
now you take those structured descriptions (not the raw videos, the text descriptions of what's IN the videos) and paste them into astra's 1,050,000 token context window. this is where the magic happens.
astra can now "see" 50 videos at once through their descriptions. it knows what every first frame looks like, what every hook says, how many scene changes each video has, where the CTA appears, what the transcript says, everything. and it can compare all of them simultaneously.
A. the video autopsy
once your pipeline has processed a video, feed the structured report into astra with this prompt:
1"here is a structured breakdown of a UGC video from our campaign. an agent extracted the frames, transcript, and on-screen text automatically.23[paste the full structured video report from the pipeline]45metrics: [views, likes, comments, shares, saves]6platform: [TikTok / Instagram Reels]7creator follower count: [X]8campaign brief: [paste brief or summarize]910analyze this video across these dimensions:1112HOOK (0-1 second):13- based on the first frame description, what visual hook is present? what emotion does it trigger?14- what is the text hook? does it create a knowledge gap, fear of missing out, or pattern interrupt?15- based on the frame sequence, would a viewer stop scrolling? why?1617RETENTION (1-5 seconds):18- does the content deliver on the hook's promise by frame 3 or does it stall?19- how many visual changes happen in the first 5 seconds?20- is there a reason to keep watching past second 3?2122STRUCTURE:23- what content format is this? (testimonial, tutorial, day-in-my-life, reaction, before/after, storytime, comparison)24- what is the information architecture? (problem-solution, list, narrative, demonstration)25- how many distinct scenes or visual changes are there across all frames?26- at what timestamp does the app first appear on screen?2728TRANSCRIPT ANALYSIS:29- does the spoken content match the brief's key messaging?30- what specific phrases does the creator use to describe the product?31- are there any claims that weren't in the brief?3233CREATOR FIT:34- based on the visual descriptions, does the creator look authentic?35- does the setting match their usual content style?3637VERDICT:38- score this video 1-5 on hook strength, retention structure, authenticity, and brief alignment39- what is the single biggest thing that would improve this video?40- would you scale this creator based on this video? why or why not?41- write a 3-line feedback message for this creator (one positive, one fix, one suggestion)"
the key difference from doing this manually: the agent watches the video, not you. the structured report captures details a human reviewer misses because it reads every frame and transcribes every word. and the feedback output is specific enough to send directly to the creator.
why this matters: we review hundreds of videos per week across 62 campaigns. the bottleneck was never finding bad videos. it was articulating WHY a video underperformed in a way a creator can act on. "make it more engaging" is useless feedback. "your hook creates a knowledge gap but the payoff comes at second 8 instead of second 2, which is why your 3-second retention drops to 22%" is feedback a creator can fix in their next draft.
this prompt gives you that level of specificity for every single video. we used to spend 15-20 minutes writing feedback per video. now we spend 2 minutes confirming what the model found.
B. side-by-side video comparison
take your best performing video and your worst from the same campaign. extract frames from both.
1"i'm going to show you two UGC videos from the same campaign for the same app. one performed well, one didn't. both had the same brief.23VIDEO A (winner):4frames: [attach 6-10 screenshots]5caption: [paste]6on-screen text: [paste]7metrics: [views, likes, comments, shares]89VIDEO B (underperformer):10frames: [attach 6-10 screenshots]11caption: [paste]12on-screen text: [paste]13metrics: [views, likes, comments, shares]1415both videos had this brief: [paste brief or summarize the 5 key components]1617analyze the gap:181. what did video A execute from the brief that video B missed?192. compare the first frame of each. which one stops the scroll and why?203. compare the hook text. which one creates a stronger knowledge gap?214. at what second does each video 'earn the next second'? identify the exact moment225. compare pacing: how many visual changes per 10 seconds in each?236. compare authenticity: which creator feels more natural with the product?247. if you could only change ONE thing about video B to match video A's structure, what would it be?2526output as a comparison table with scores for each dimension, then a 3-sentence action item i can send directly to video B's creator as feedback."
why this matters: the comparison is where the real learning happens. when you look at one video in isolation, everything feels subjective. when you put the winner and loser next to each other with the same brief, the structural difference becomes obvious. the 3-sentence feedback output is deliberate. creators don't read long emails. they need one specific thing to fix.
C. batch format analysis across competitors
this one requires more setup but it's the most valuable research prompt in this entire article.
go to 3 competitor apps' TikTok or IG accounts. screenshot the first frame of their 10 most recent videos (30 screenshots total). note the view count for each.
1"i'm analyzing the UGC strategy of 3 competitor apps in my category. for each app i've captured the first frame of their 10 most recent videos along with the view count.23COMPETITOR A: [app name]4[attach 10 first-frame screenshots with view counts listed]56COMPETITOR B: [app name]7[attach 10 first-frame screenshots with view counts listed]89COMPETITOR C: [app name]10[attach 10 first-frame screenshots with view counts listed]1112MY APP: [app name]13[attach 10 first-frame screenshots with view counts listed]1415analyze across all 40 videos:16171. VISUAL PATTERNS: what do the top 5 highest-performing first frames have in common? (face position, text placement, colors, props, background, lighting)182. TEXT HOOK PATTERNS: group all 40 on-screen text hooks by structure type. which structure type has the highest median views?193. FORMAT DISTRIBUTION: what % of each competitor's content is testimonial vs tutorial vs reaction vs other? which format type wins?204. WHAT MY COMPETITORS DO THAT I DON'T: identify any visual or structural pattern that appears in 2+ competitor accounts but not in mine215. WHAT I DO THAT NO ONE ELSE DOES: identify anything unique to my content that could be a differentiator or a blind spot226. THE PLAY: based on all 40 videos, what is the single highest-opportunity content format and hook structure that i should test next week? be specific - give me the exact first-frame composition, text hook structure, and content flow."
why this matters: this is competitive intelligence that used to require hiring an analyst for a week. you're getting a structural breakdown of your entire competitive landscape from the visual layer. the key insight for us was discovering that 3 of our competitor apps were using the exact same first-frame template (text in top-left, face in bottom-right, colored background) and we weren't. we tested it the next week and our hook rate jumped 18%.
PART 2: CAMPAIGN INTELLIGENCE (prompts 4-8)
this is where Astra's 1,050,000 token context window stops being a spec sheet and starts being a weapon.
for context: our entire database of 22,504 video captions with metrics fits in that window with room to spare. for the first time ever, the whole book fits on the desk instead of one page at a time.
D. caption clustering
this is prompt #1 in order of impact. if you only run one prompt from this entire article, run this one.
1"here are all the captions and on-screen hooks from our last 30 days of UGC content, with the view count for each.23[paste: one line per video. format: 'HOOK TEXT | CAPTION | VIEWS | PLATFORM | CREATOR']45do not cluster these by topic. cluster them by STRUCTURE.67a structure is the skeleton of the hook, not what it's about. examples:8- 'POV: you [action] and [unexpected result]' is one structure9- '[number] things i wish i knew about [topic]' is another structure10- 'i can't believe nobody told me about [thing]' is another11- 'this is what i use instead of [alternative]' is another1213find every distinct hook structure in this dataset. for each structure:141. name it with a short label (e.g., 'POV + unexpected result')152. count how many videos used it163. calculate the median views for that structure174. calculate the average views (so i can see if one outlier is inflating it)185. list the top 3 and bottom 3 performers within that cluster1920then rank all structures by median views, highest first.2122finally: compare the distribution of structures in my brief against the distribution of structures in actual posts. are creators following the brief's intended structure or have they drifted into their own patterns?2324output as a table. then give me a 3-sentence recommendation: which structure should i double down on, which should i kill, and which structure from a competitor should i test."
why this matters: this is the prompt that found our 4.2x problem. we had two campaigns for the same app. identical briefs. campaign A did 12,333 views per post. campaign B did 2,921. campaign B had more posts and more creators.
the caption clustering showed that campaign A had converged on one hook structure and repeated it across all creators. campaign B had 41 creators each running their own interpretation. the brief was the same. the execution drifted. nothing in our process caught it because no human reads 3,109 captions side by side.
a 1,050,000 token window does. and now we run this every monday.
E. brief drift detection
your brief has 5 components. your creators execute maybe 3 of them. the other 2 disappear silently and you don't notice until the campaign wraps.
1"here is our creative brief for this campaign:23[paste your full brief - objective, key messaging, audience insight, content requirements, creative inspiration]45and here are the last 200 captions + on-screen hooks from creators on this campaign:67[paste: one line per video. format: 'CREATOR | HOOK TEXT | CAPTION | VIEWS']89for each of the 5 brief components, score every video on whether it executed that component:10- objective alignment: does the video point toward the campaign goal? (yes/partial/no)11- key messaging: does the video use or paraphrase the messaging points? (yes/partial/no)12- audience insight: does the video speak to the target persona? (yes/partial/no)13- content requirements: does the video follow format/length/CTA requirements? (yes/partial/no)14- creative inspiration: does the video reflect the examples and mood we provided? (yes/partial/no)1516then group by creator. for each creator, show their compliance rate per brief component.1718identify:191. which brief component has the lowest compliance rate overall?202. which creators consistently miss the same component?213. is there a correlation between brief compliance and views? (sometimes the 'off-brief' videos win - i need to know that too)224. for creators scoring below 60% compliance: write me a specific 2-sentence message i can send them to course-correct without killing their creativity."
why this matters: the messaging block is the one that silently disappears. we proved it. across 41 creators on one campaign, only 7 consistently executed the key messaging. the other 34 were making content that looked right but said the wrong thing. the course-correction messages alone saved that campaign $40,000 in wasted creator fees the following month.
F. comment mining
this one is embarrassing. we had 131,182 comments across our campaigns and nobody was reading them systematically until 3 months ago.
1"here are the last 5,000 comments from our campaign's videos.23[paste comments - one per line, include the video URL or creator name if possible]45mine these comments for:671. PRODUCT LANGUAGE: what exact words and phrases do commenters use to describe our app? not our marketing language - their language. list the top 20 phrases by frequency.892. OBJECTIONS: what concerns, doubts, or reasons NOT to download appear? group them by theme. for each theme, write a hook that directly addresses that objection.10113. USE CASES WE DIDN'T BRIEF: are commenters describing uses for the app that we never mentioned in any brief? list them. these are content angles we're missing.12134. COMPETITOR MENTIONS: do any comments mention competing apps by name? what do they say? what can we learn from why they're comparing?14155. VIRAL COMMENT HOOKS: which comments got the most likes? what did they say? comments with 100+ likes often contain the exact hook phrasing that would work as a video hook.16176. CREATOR-SPECIFIC PATTERNS: do certain creators attract different types of comments? (e.g., one creator gets mostly 'where do i download' while another gets 'is this even real'). what does that tell us about which creators drive conversions vs which drive skepticism?1819output each section as a separate table. then write me 10 new video hooks based entirely on the language, objections, and use cases you found in these comments. these hooks should sound like a real person talking, not a marketer - because they literally come from real people."
why this matters: your next viral hook is already written. it's sitting in your comment section in the exact words your audience uses to describe your product. the phrase "wait this actually works?" appeared 847 times across our campaigns. that became a hook. it did 2.1M views because it was already validated language from the audience itself.
G. creator performance analysis
stop making creator decisions based on follower count. this prompt builds a scoring model from your own data.
1"here is the performance data for every creator in our program:23[paste: one row per creator. format: 'CREATOR | FOLLOWERS | TOTAL POSTS | TOTAL VIEWS | AVG VIEWS/POST | TOP VIDEO VIEWS | BRIEF COMPLIANCE % | MONTHS ACTIVE | COMPENSATION/MONTH']45build me a creator scoring model based on this data. do not use follower count as a factor.67the model should score each creator on:81. consistency: standard deviation of views across their posts (lower = more reliable)92. ceiling: their top single video performance relative to the campaign average103. efficiency: views per dollar of compensation114. trend: are their last 10 posts trending up or down vs their first 10?125. hit rate: what % of their posts exceed the campaign median?1314tier every creator into S/A/B/C:15- S tier: top 10% on composite score. these get increased budget and creative freedom.16- A tier: top 25%. reliable performers. keep current terms.17- B tier: middle 50%. evaluate on a per-creator basis.18- C tier: bottom 25%. flag for potential replacement.1920for each C-tier creator, tell me:21- what specifically is underperforming (consistency? ceiling? efficiency?)22- is there evidence they could improve with a brief adjustment, or is this a structural mismatch?23- replacement recommendation: keep with intervention, pause, or drop?2425output the full ranked table, then a summary: how many creators should i scale up, how many need intervention, and how many should i replace this month?"
why this matters: we had a creator with 4,400 followers who hit 28M views across a campaign. we had another with 85,000 followers who averaged 306 views per post. follower count told us the opposite of reality. this scoring model finds the creators who actually perform, and it updates every week as new data comes in. the efficiency metric alone saved us $12,000 last quarter by catching overpaid underperformers.
H. cross-platform format translation
what works on TikTok doesn't automatically work on Instagram Reels. the mechanics are different. this prompt adapts your winners.
1"here is our top performing TikTok video:23hook: [paste on-screen text]4caption: [paste full caption]5format: [describe: talking head, slideshow, screen recording, etc.]6length: [X seconds]7metrics on TikTok: [views, likes, comments, shares]89and here is our performance data comparing TikTok vs Instagram Reels across our campaign:1011TikTok avg views/post: [X]12IG Reels avg views/post: [X]13TikTok hook rate benchmark: >60%14IG Reels hook rate benchmark: >50%15TikTok avg video length that performs: [X seconds]16IG Reels avg video length that performs: [X seconds]1718translate this winning TikTok format for Instagram Reels. specifically:191. rewrite the hook for IG's different scroll behavior (IG users scroll slower, hooks can be slightly less aggressive)202. adjust the pacing - IG Reels tend to reward slightly longer establishing shots213. suggest caption adjustments (IG captions are more visible than TikTok's)224. recommend any visual changes (IG favors slightly higher production quality)235. write the full adapted brief i can send to a creator for the IG version2425do the same in reverse: take our top IG Reel and translate it for TikTok."
why this matters: we tracked this across 122,504 posts. Instagram Reels delivered 3.2x more views per post than TikTok in our data. but the formats that won on each platform were structurally different. the same video cross-posted without adaptation performed 40-60% worse than a platform-native version. this prompt gives you the platform-native brief without having to figure out the differences yourself.
PART 3: BRIEF GENERATION (prompts 9-12)
most people use AI to write briefs from scratch. that gives you generic output because the model is writing from what it learned on the internet, not from what works for your product.
these prompts write briefs from your own data.
I. hook generation from your own winners
this is the difference between "generate me 30 hooks" and "generate me 30 hooks in the exact structural pattern that already works for my app."
1"here are my 50 highest-performing video hooks from the last 90 days, ranked by views:23[paste: one per line. format: 'HOOK TEXT | VIEWS | PLATFORM']45first, identify the 3-4 structural patterns these hooks share. name each pattern.67then generate 30 new hooks following these rules:81. every hook must follow one of the identified structural patterns92. no hook should be a direct copy - they should feel fresh while being structurally identical103. vary the specificity: some should name the app, some should be product-agnostic (for ghost accounts)114. include 5 hooks specifically designed for the 'POV:' format125. include 5 hooks designed for the 'list' format (3 things, 5 reasons, etc.)136. include 5 hooks designed for the 'before/after' or 'transformation' format147. for each hook, rate its predicted performance (high/medium/low) based on how closely it matches the winning patterns1516output in a table with columns: hook text, structure pattern, format type, predicted performance, suggested creator type (face-to-camera vs voiceover vs faceless).1718then pick the top 5 hooks you think will outperform and explain why each one works structurally."
why this matters: when you prompt for hooks without feeding the model your own data, you get the same hooks every other brand is running. when you feed it your winners first, every output is calibrated to what already performs for YOUR audience. the difference is the gap between a creator script that sounds like ChatGPT and one that sounds like it was written by someone who watches your content daily.
J. personalized brief per creator
this is the one that blew our operations team's mind. instead of one brief for all creators, you generate a brief adapted to each creator's style.
1"here is our campaign brief:23[paste full brief with all 5 components]45and here is creator data for [creator name]:67account: [@handle8]9platform: [TikTok / IG]10follower count: [X]11content style: [describe their typical video style]12their 3 best performing videos: [paste hooks + view counts]13their 3 worst performing videos: [paste hooks + view counts]14what makes their content unique: [1-2 sentences]15past campaign performance with us: [paste if available]1617generate a personalized version of the campaign brief specifically for this creator.1819the personalized brief should:201. keep all 5 brief components intact - do not remove any requirement212. translate the messaging into language that matches this creator's natural voice223. suggest a specific hook from their winning patterns that fits our campaign objective234. recommend a content format based on what this creator does best (not what we usually ask for)245. flag any potential friction points - things in our brief that this creator might struggle with based on their past content256. be under 300 words. creators don't read long briefs.2627write it in second person ('you') and make it sound like a creative director talking to a collaborator, not a brand talking to a vendor."
why this matters: we tested this across one campaign. 20 creators got the standard brief. 20 got the personalized version. the personalized-brief group produced videos with 31% higher brief compliance and 22% higher average views. same app, same campaign, same month. the only difference was that each creator received instructions translated into their own creative language instead of corporate marketing speak. at scale across 590 creators, this alone is worth more than any other prompt in this list.
K. campaign script writer
for creators who need more direction than a brief. some want the exact script. this writes it from your data, not from thin air.
1"i need a video script for a [X second] TikTok/Reel promoting [app name].23the script should follow this proven hook structure from our campaign: [paste your winning hook pattern]45here is the campaign brief: [paste brief]6here is the creator who will film it: [paste their style description and best hooks]7here is our target audience: [describe]89write the full script:1011SECOND 0-1 (HOOK):12- on-screen text: [exact text]13- what the creator says: [exact dialogue]14- what the viewer sees: [visual direction]1516SECOND 1-5 (RETENTION):17- on-screen text: [exact text]18- dialogue: [exact words]19- visual: [what happens on screen]2021SECOND 5-15 (VALUE):22- break into 2-3 scene changes23- each scene: text + dialogue + visual2425SECOND 15-END (CTA):26- on-screen text: [exact text]27- dialogue: [exact words]28- visual: [what happens]2930the script should feel like something this specific creator would naturally say. not corporate. not scripted-sounding. like they're talking to a friend.3132also provide:33- 3 alternative hooks for A/B testing34- suggested filming location (based on the creator's typical content)35- one 'pattern interrupt' moment to prevent mid-video drop-off"
why this matters: we use the "guide, don't script" approach as default. but some creators need more structure, especially newer ones. the difference between a script written from your campaign data vs a script written from nothing is the difference between a video that sounds native and one that sounds like an ad. this prompt produces scripts that creators actually want to film because they see their own voice in the words.
M. weekly research brief
replace your team's 2-3 hour research ritual with a 15-minute prompt session.
1"here is our campaign performance from the last 7 days:23[paste: one row per video. format: 'DATE | CREATOR | PLATFORM | HOOK | VIEWS | LIKES | COMMENTS | SHARES']45and here is the performance from the previous 7 days for comparison:67[paste same format]89run the weekly research analysis:10111. WEEK OVER WEEK: did total views go up or down? by how much? which creators drove the change?122. HOOK PERFORMANCE: which hooks from this week outperformed the campaign average? which underperformed? what structural pattern do the winners share?133. PLATFORM SPLIT: how did TikTok vs IG Reels perform this week compared to last? is one platform trending up while the other flattens?144. CREATOR HEALTH: flag any creator whose average views dropped more than 30% week over week. possible causes?155. FORMAT TESTS: if we ran any new formats this week, how did they compare to our proven formats?166. BRIEF UPDATES: based on this week's data, should we update anything in the active brief? be specific.1718end with:19- 3 things to double down on next week20- 2 things to stop doing21- 1 experiment to test2223keep the entire output under 500 words. this goes in our monday standup."
why this matters: the monday research ritual was 2-3 hours of manual spreadsheet analysis. this prompt does it in 15 minutes and catches patterns a human misses because it reads every single row instead of sampling. the "1 experiment to test" output is what keeps campaigns from going stale. we've been running this weekly for 8 campaigns and the experiment suggestions have a 40% hit rate. that's 40% of the time the model surfaces a test we wouldn't have thought of that actually works.
PART 4: OPERATIONS (prompts 13-15)
L. creator feedback generator
the hardest operational task in UGC is writing feedback that is specific enough to change behavior without killing the creator's motivation.
1"here is a draft video from a creator:23creator: [@handle4]5frames: [attach screenshots if available, or describe the video]6hook text: [paste]7caption: [paste]8video length: [X seconds]9brief they were working from: [paste brief]1011and here is our scoring rubric:12- hook strength (1-5): does the first second stop the scroll?13- retention structure (1-5): does the pacing keep viewers past 3 seconds?14- brief alignment (1-5): does the content match what we asked for?15- authenticity (1-5): does this feel like the creator's real content or an ad?16- technical quality (1-5): lighting, audio, framing acceptable?1718score this video on all 5 dimensions.1920then write feedback in exactly this format:21LINE 1: one specific thing they did well (always lead positive)22LINE 2: one specific thing to change before posting (the single highest-impact fix)23LINE 3: one creative suggestion for their next video (plant a seed, don't mandate)2425the feedback must be:26- under 50 words total27- written in casual, encouraging tone28- specific enough that the creator knows exactly what to do differently29- never use the words 'engaging,' 'compelling,' or 'great job'"
why this matters: we review hundreds of drafts per week. consistency is the killer. one UGC manager writes a paragraph of feedback, another writes "looks good." this prompt standardizes the output without standardizing the voice. the ban on "engaging" and "compelling" forces specificity. "your hook creates curiosity but the payoff comes at second 8 instead of second 3" is 100x more useful than "make the hook more engaging."
N. campaign post-mortem
run this when a campaign ends. it catches the patterns that the wrap report misses.
1"here is the complete data from a campaign that just ended:23campaign name: [X]4duration: [X weeks]5total creators: [X]6total posts: [X]7total views: [X]8average views per post: [X]9budget spent: [$X]10CPV (cost per view): [$X]1112top 10 videos: [paste with hooks, creators, views]13bottom 10 videos: [paste with hooks, creators, views]14brief used: [paste]1516run a full post-mortem:17181. WHAT WORKED: identify the structural patterns in the top 10. what do they share? (hook type, format, creator demo, platform, time of posting)192. WHAT DIDN'T: identify the structural patterns in the bottom 10. what do they share?203. THE 80/20: what % of views came from the top 20% of posts? what % came from the top 3 creators?214. BRIEF EFFECTIVENESS: based on the data, which brief components correlated with high performance? which were ignored without consequence?225. CREATOR TIERS: which creators should be promoted to the next campaign? which should be dropped?236. THE MISSED OPPORTUNITY: what format, hook style, or platform did we NOT test that the data suggests we should have?2425end with a 5-point brief revision for the next campaign based on what this data shows. not what we think should work. what the data says works."
why this matters: post-mortems are the most important document in a campaign and the one most people skip. this prompt means there's no excuse. the "missed opportunity" section is worth the entire exercise. it consistently surfaces the obvious-in-hindsight test that nobody thought to run.
O. the full campaign audit
this is the nuclear option. dump everything into the context window and ask the question nobody asks.
1"i'm going to give you our complete campaign data. this includes every video, every caption, every creator, every metric, and our brief. this is [X] rows of data representing [X] months of work.23[paste the full dataset]45before you start analyzing, confirm: how many rows did you receive? how many unique creators? what date range does this cover? i want to make sure nothing was truncated.67then answer the 5 questions that determine whether this campaign was actually good or just felt good:891. CONCENTRATION RISK: what % of total views came from the top 3 creators? if more than 60%, this campaign is fragile. one creator leaving kills it.10112. HOOK CONVERGENCE: did the campaign converge on a winning hook structure over time, or did hook styles stay scattered? convergence = good (the system is learning). scatter = bad (every creator is running solo).12133. VELOCITY TREND: are views per post trending up or down over the campaign's lifetime? plot the weekly averages. a campaign that starts hot and fades is different from one that builds momentum.14154. THE REAL CPV: not the headline CPV. calculate CPV excluding the top 3 outlier videos. that's the performance you can actually rely on and scale.16175. REPEATABILITY: if i launched this exact campaign again tomorrow with 20 new creators and the same brief, what would the expected views per post be? high variance = lucky, not good. low variance = a system that works.1819be honest. if the campaign was mediocre, say so. the only thing worse than a failed campaign is one that felt successful but can't be repeated."
why this matters: i'll tell you a number that should terrify anyone running UGC at scale. in our data, 3 creators out of 590 account for 38% of all views. that's not a program. that's three people. this prompt catches that reality and forces you to build around it instead of pretending your averages represent your actual performance.
how to use these prompts
do not run all 15 at once. here's the order.
day 1: load your campaign context. then run prompt 4 (caption clustering). this is the single fastest insight you will get. 30 minutes of work. do it today.
day 2-3: run prompt 1 (video autopsy) on your best and worst video. then run prompt 2 (side-by-side comparison). you now understand WHY your content performs the way it does.
week 1: run prompts 5 (brief drift) and 6 (comment mining). you now know what your brief is actually producing vs what you intended, and what your audience is actually saying.
week 2: run prompts 7 (creator scoring) and 9 (hook generation). you now know which creators to scale and have 30 new hooks calibrated to your data.
week 3: run prompt 10 (personalized briefs) for your top 10 creators. run prompt 12 (weekly research) every monday from this point forward.
ongoing: prompt 13 (creator feedback) on every draft. prompt 14 (post-mortem) at every campaign end. prompt 15 (full audit) once per quarter.
90 days of running this system and you will have better campaign intelligence than agencies that have been operating for years. i know because we built these prompts from managing 2.6 billion views worth of campaigns. the prompts are the system. the system is what scales.
the real talk
95% of people reading this will save it and never paste a single prompt. that's fine.
if you want my team to run this entire system for your app (every campaign, every creator, every brief, every week of execution) that's what we do at The Viral App.
2.6 billion organic views. 590 creators. 62 campaigns. zero ad spend.
