19 Creative use cases · newest Nov 11, 2026 · all 87 →
AI agent use cases for ad creative.
Real ad creative workflows people are running with AI agents, newest first. Each one links to the original post, lists the tools used, and includes the prompt where it was shared.
Matthew Berman describes a system he packaged as the "Ad Spy Kit": seven Claude skills (ad-spy-kit to run it all, resolve, pull, winners, analyze, counter, render). You give it 3 to 5 competitor brands by name, IG handle or Ad Library link. It pulls their live Meta ads through StealAds or Scrape Creators, sorts them by days running and number of versions, and labels winners (60–120+ days with 3+ variations) versus tests and flops. It then pulls hooks and offers word for word from the transcript, on-screen text and caption, with each quote tied to its source ad so you can check it, maps the promises, formats and ignored objections, finds what every brand is saying and which formats are crowded, and outputs a board with a headline, hook, what to show and the offer built around the gap nobody has taken. He frames it as replacing a $149/mo Foreplay plan plus an $8K/mo creative strategist.
Claude skillsScrape CreatorsStealAdsMeta Ad Library
Evan Swanson, who manages brands doing $60M a year, keeps an "operator OS" folder with his operating principles and brand guidelines and opens it in Claude Code or ChatGPT Codex, so he can switch AIs without losing context; the folder is backed up to GitHub. He connects MCPs instead of all-in-one tools, including BrandSearch (competitor traffic, landing pages, social and Meta ads), Higgsfield (image and video generation), Shopify, Canva, Gmail and Slack, plus Amazon research and PPC tools. One of his prompts: use the BrandSearch MCP to scan competitor websites for their top-performing ads and landing pages, then use the brand guidelines in the folder and Higgsfield to generate similar static ads. Others review PPC performance against a 20% margin target and estimate cost per acquisition from keyword CPC and conversion rates.
Claude CodeCodexBrandSearch MCPHiggsfield MCPShopify MCPGitHub
Use the brand search MCP to scan competitor websites for their top-performing ads and landing pages. Then use my brand guidelines and Higgsfield to generate similar static ads according to our guidelines in the folder
@pounddz shared a one-person AI UGC workflow: use Gethookd to scan competitors' best-performing ads and pick a format and copy structure that fits your product, have Claude rewrite that copy using your own customer research, then have Claude generate the images and the video (paired with Google Omni), cut the awkward parts and add on-screen text. He says he dictated the whole brief by voice and one-shotted the video and copy. Useful for brands that want cheap volume of UGC-style concepts built on proven formats, but keep a human pass on claims and on how close the copy sits to the original ad.
Here are a competitor's top-performing ads from Gethookd and my customer research. Pick the format that fits my product, rewrite the copy in my customers' words, then write the UGC script, generate the scenes, and list the on-screen text for each cut.
Roshan (@roshanp), whose team publishes about 1,000 Meta ads a month for booked-call offers, shared 50 tactics built around an AI agent (Grok Bot). The agent work: build an ad scripting guide from past winners and losers; break the top 3-4 ads into angle, messaging, offer framing, who it calls out and proof; pull the exact words buyers use on sales calls each month; diagnose every ad that stops working (no spend = weak angle, low hook rate = weak first 3 seconds, high hook but low hold = weak body, high hold but low CTR = weak CTA or offer, clicks but no conversions = ad and page don't match); and every Monday pull CPM, unique outbound CTR, hook rate, show rate, qualified rate and close rate for every ad, then pick the one metric to improve that week. His numbers are self-reported for lead-gen accounts (for example a 5-10% average hit rate, with a winner defined as an ad that spends 5x target cost per result and stays inside KPI); for DTC, swap calls for purchases. The post links an article with the 12 prompts behind the system. The article's setup step: connect Meta Ads read-only, load a business-context document (offer, audience, funnel, targets, provable results only), and have the agent send any budget, status or targeting change to you as a draft with the current setting, the new one and why.
Pull the transcripts of my [10] ads with the lowest cost per qualified call and the [10] with the highest, and write me a practical ad scripting guide.
Josh Jackson (@IamJoshJackson), who sells ad creative to brands, shared a five-step outbound flow he's testing with the @gimmegimmeok MCP and says is "10x'ing" his outbound (self-reported, no numbers). The agent: 1) looks at his recent LinkedIn connections and picks the ones in his ideal customer profile, 2) finds the brands they work with, 3) identifies a brand that's running new Meta ads, 4) makes 4 new Meta ads for that brand, and 5) sends an email with the ads and asks to chat. It's useful for agencies, freelancers and creative studios pitching DTC brands with finished work instead of a deck. Review the ads and the email yourself before anything is sent, and use only the brand's public ads and assets.
Video Factory (MIT licensed, by LordVayron) is a repo that Claude Code or Codex installs from one message, then asks five onboarding questions: what you make, where you post, language and tone, your brand or characters, and budget per video. For DTC it makes 15–30 second 9:16 UGC-style product ads from your product shots (the README lists $1–3 typical API cost) and 30–35 second launch films from your landing page ($4–8). You approve the concept, script, cast and a 'Production Book' with the cost before anything is generated; after that the budget is locked in code and any job that would exceed it is refused. Voices are recorded first so shot lengths are fixed before paying for video, and every take is checked for face drift, unplanned cuts and garbled on-screen text. It needs APIMart or Kie.ai plus ElevenLabs keys, kept in a local .env file. The README asks you to label AI-generated ads where platforms require it and not to use real people or brands you don't own.
Marek Dąbrowski's creative-strategist skill has Claude read the account through the Meta Ads MCP (spend, frequency, CTR, CPM, conversions and CPA per ad, plus your target CPA) and name any number that's missing. Step 1: group ads into concepts, meaning ads that share the same promise and proof (a video and a static with the same promise are one concept), and keep that ad-to-concept table so next week compares like with like. A concept only gets a status after spending at least 3x target CPA with 10+ conversions. Step 2: compare each concept with its own last 30 days. Winner: CPA at or below target and CTR within 20% of baseline. Tired winner: at target for 14+ days, then frequency up 50%+, CTR down 30%+ and CPA up 20%+ at once, with CPM within 15% so the auction is ruled out. Watchlist: two of those signs, or a decline under 7 days old. Replace: tired for 14+ days with no recovery at stable spend (pausing stays your call). Never worked: never within 20% of target after 3x target CPA. Frequency alone never makes a concept tired. Step 3: one table with concept, status, evidence, likely issue, replacement angle and confidence; if CPM rose while CTR held, it flags auction pressure instead of creative.
Ferdinand Terme built an MCP-compatible "brand brain" for Byredo so Claude stops starting from scratch each time. He gave it byredo.com plus five campaign pages, screenshots of three of the brand's best-performing Meta ads and packshots of core products. It learned the logo, colors, typeface, 23 products and their ranges, how the brand writes, and the formula behind the top Meta ads (huge stencil wordmark on top, one hero, one small caption, zero sales talk). From there he worked by chat: a Black Saffron static "in the style of our best Meta ads", then animate it, resize it for every placement, give four other angles, and a clean plate without copy. Output: a 9:16 static on formula, a 6-second video of it, 1:1, 4:5, 9:16 and 16:9 versions in one run, and four new angles (product hero, bold statement, skin close-up, dressing table at night). He used Pletor but says the steps work with any platform.
Make a Black Saffron static in the style of our best Meta ads.
Animate it.
Resize it for every placement.
Give me 4 other creative angles.
Remove the copy, I want a clean plate.
Mike Futia vibe-coded a review scraper with Claude. You paste a Shopify product URL, yours or a competitor's, and it tags the reviews as your brand or a competitor, scrapes all of them and saves them into the brand's research workspace. Then you chat with the reviews to pull out buying motivations, recurring objections and unmet needs. Out come customer phrases for hooks and copy, objections your ads need to answer, gaps in competitor products to test as angles, and ad concepts with hooks, primary text and suggested static visuals that you keep refining in the same chat. Only scrape reviews that are publicly shown on the product page, and check a sample of quotes by hand before they go into ads.
Pull 20 to 50 real reviews, mostly 3 and 4 stars, from your store, Amazon, a competitor's listing, Reddit or the comments on your ads, copied word for word. Ask Claude or ChatGPT to quote, not summarize: for each review, pull out the moment they bought for, what they hoped would happen, what let them down and any product they compared you to, then group them into the 5 most common themes. Ask for 3 hooks per theme in the customer's own words (one naming the moment, one the fear, one answering the objection), delete the half that sound like AI, and test the themes as separate concepts. Check theme counts against the real reviews, and never turn a customer's words into a claim you can't back up. The author sells a Meta ads prompt pack, but the workflow needs only a chat window.
Here are real customer reviews. Don't summarize them. For each review, pull out: (1) the exact moment or situation they bought for, (2) what they hoped would happen, (3) what disappointed or worried them, (4) any product they compared us to. Quote their words. Then group the results into the 5 most repeated themes, ranked by how often they appear.
Riley Timmins of Kintsugi says a set of image ads used to cost about $2,000 and take a week. His team built 10 Meta ads in about 2 hours on a Claude subscription. His rule is to start with basic image creatives and prove they work before asking for more budget. It's a low-cost way to get a first creative test live: draft 10 simple static ads with Claude, launch them, and use the results to justify spending on bigger production.
Pull your niche's long-running ads into Claude through a Meta Ad Library MCP connector (or a script Claude writes) and sort every competitor ad active for 14+ days by visual format: drama skit, raw UGC, problem b-roll and founder story. Then have Claude, acting as a direct-response creative director, write four concepts in different formats, each aimed at a different buyer pain point, with scene-by-scene visual and voice prompts. Generate the scenes with your video model, put them in one folder, and have Claude Code with local FFmpeg (or a video MCP) assemble the four concepts and make five pacing and hook variations of each, with captions burned in. You end up with 20 ads that differ in concept and format, not just in hook.
Act as a direct-response creative director. Take my product and write 4 completely different ad concepts targeting 4 different buyer pain points: 1 long-form drama ad, 1 hyper-real UGC testimonial, 1 contrarian problem/solution ad, and 1 visual demonstration. Include timestamped scene-by-scene visual and voice prompts for each
A Meta ads operator had Claude build each client an internal dashboard with four parts. 1) Main KPIs, costs and creative win rate, which shows whether the process works or you're getting lucky. 2) A Kanban pipeline from concept to copy to production to live, paused or winner, so bottlenecks are obvious. 3) A batch view: each idea becomes 5 or 6 creatives with different hooks or visuals, so you can tell if the idea won or only one hook did. 4) Landing page performance with a map of which ads send traffic to which page, because "half of bad ads" are an ad-to-page mismatch. It started in Airtable, but Claude's permissions there were limited, so they had Claude convert it to HTML and host it on Vercel behind a password.
With a UGC tracking tool connected to Claude over MCP, one team analyzed 18,736 September UGC videos by talking to it, replacing what they say was 90% of their day in spreadsheets. It ranks hooks, formats and angles by views and consistency, says which creators to scale or coach, turns the top videos into 10 new concepts with the winners that inspired each one, and lists unanswered comment questions per creator. The author co-founded Trackr and measures app downloads, but the same prompts work on a DTC brand's UGC and Spark Ads library.
Analyze every video from [CAMPAIGN] over [TIME PERIOD]. Identify the top-performing hooks, formats, and messaging angles, ranking them by average views, engagement, and consistency. Show me the specific videos supporting each finding.
A Monday research run sweeps 15 competitors in the Meta Ad Library (flagging ads live 60+ days), tears down their landing pages and mines Reddit language. It then writes 25 hooks and 5 storyboards and produces 10 to 15 videos a week in Higgsfield. A daily Meta check flags frequency over 3.5 or falling ROAS, and a weekly kill and scale list feeds the next round.
A growth marketer with no coding background asked Claude Code to build a Figma plugin in about an hour. He picks the creative frame, pastes headline copy once from a Google Sheet, and the plugin generates every permutation of that ad in one click, saving him nearly 30 minutes each time he updates a big batch of creatives across aspect ratios.
Claude, I'm working in Figma. I really want to be able to solve this challenge of this repetitive copy and pasting. Can you help me build a Figma plugin to help me solve my challenge?
Runs weekly to scrape competitors' active ads from the Meta Ad Library, summarize recurring angles and hooks, and append a creative brief to Google Sheets.
A four-workflow system that collects competitor Facebook ads, scores them by run time and variation count, analyzes visuals and hooks, and generates on-brand copy and new ad images, all saved to Sheets and Drive.
Refresh customer intel, pull competitor ads from the Ad Library, diagnose dying creatives, generate on-brand replacements, animate winners, apply Meta's AI content label, and push live. The author says it takes about 30 minutes per ad account.