The Ultimate Guide to AI-Driven Creatives

Most performance teams describe the shift to AI-driven creatives as a production upgrade. “We can generate fifty variants in an afternoon.” That framing is right about what’s happening. It’s wrong about why it matters.

The promise of AI-driven creatives isn’t that you ship more ads. It’s that the winning teams test faster, learn faster. A traditional UGC cycle (concept brief, casting, talent release, shoot direction, edit, publish) runs seven to fourteen days. With AI in the production layer, the same cycle runs in one workday or less. That changes which questions you can ask the data.

When the loop was fourteen days, you tested one or two big bets per cycle and prayed. When the loop is hours, you can test the bets, the variations on the bets, and the variations on the variations. The question stops being “did this concept work” and becomes “which of the eight subcomponents inside this concept did the work.”

That’s the leverage. And almost every team chasing AI creatives misses it, because they default to one of two failure modes. They ship sloppy AI fast, or they ship beautiful AI that quietly underperforms. Both failures share the same root cause: they treat AI as a strategy tool instead of a production tool, and pull humans out of the loop right when the loop needs them most.

This guide covers what stays human, what stays AI, and the chapter-by-chapter playbook for shipping AI-driven creatives that perform on Meta and TikTok in 2026.

Three trends that explain where the leverage actually is

Across the subscription-app and gaming creative work we’ve seen this year, three trends keep showing up. They explain why AI-driven workflows are unlocking step-changes for some teams, while making other teams ship more bad ads faster.

Trend 1: Audience diversity beats audience monoculture. Default creative targets the easiest-to-imagine user. Reviews, user interviews, and search data surface non-obvious ones: reluctant converts, unexpected use cases, accidental fans, switchers from adjacent categories. Teams that target micro-audiences (millennial dads, women over fifty, ESL learners, night-shift workers, parents of neurodivergent kids) find scale where the monoculture-targeted teams hit a ceiling. AI helps because per-segment creative is now cheap. The constraint is no longer production. It’s whether you’ve looked hard enough at your user base.

Trend 2: Format diversity escapes algorithmic flattening. Most apps get trapped in one format because it’s winning. All video, no static. All product mockup, no UGC. All faceless, no on-camera. Breaking format (static vs. video, UGC vs. product-shot, walls of text vs. clean hero shots) unlocks new audience segments. The algorithm reads format diversity as audience diversity. AI lets one concept ship in five formats by Friday.

Trend 3: Human in the loop is the antidote to AI slop. Everyone is generating ads with AI. Most of it is slop: generic, no editorial voice, no brand memory, no curiosity gap. Winners use AI for volume, humans for the strategy layer: what the ad says, who it’s for, what the hook must do. The teams shipping the highest-performing AI creative have grown their creative teams in the last year, not shrunk them. The human work moves from production (which AI eats) to brief writing and curation (which AI cannot).

All three trends require volume. You can’t diversify across audiences or formats if you’re shipping three ads a week. Once you hit twenty to forty per week, the three trends separate teams that learn from teams that just spend.

What this guide covers

1. What carries over from the UGC playbook
2. Why iteration speed, not production speed, is the real lever
3. Chapter 1: Briefing AI hooks (text, visual, spoken)
4. Chapter 2: Visual elements at AI scale
5. Chapter 3: Generating scripts that perform
6. Chapter 4: Iterating across the eight axes
7. Chapter 5: Safe zones
8. Chapter 6: The AI production workflow
9. Chapter 7: Humans in the loop, the drafting and curation checklist
10. Chapter 8: At-scale operational discipline
11. Campaign structures that make sense
12. FAQ

What carries over from the UGC playbook

We’ve written the Ultimate Guide to UGC for Performance Marketing over years of UGC production. When we started shifting clients to AI-driven creatives, we expected the playbook to change. Most of it didn’t.

What we see: Two of seven chapters die. Five carry over almost unchanged. One, iterations, gets ten times bigger. That last one is where the actual leverage sits.

Why iteration speed is the real lever

When the iteration loop ran in days, you tested concepts. One concept per cycle, maybe two if the team was big. The post-mortem on a losing concept was structural: “did this angle work, yes or no?” Answers came in two-week increments.

When the iteration loop runs in hours, you stop testing concepts. You start testing axes. The same concept fans out into hook copy variants, visual hook variants, format variants, music bed variants, VO variants, scene order variants, color palette variants, duration variants. The post-mortem is no longer “did this angle work?” It’s “we know the angle works. Which combination of layers maximizes its lift?”

That changes which decisions are bottlenecks. Production becomes near-free. Strategy and curation become the constraints. The teams that miss this shift make one of two failures, both rooted in inverting where AI plugs in.

The first failure is sloppy AI. Generate fast, ship faster, skip the testing structure. The variant volume goes up but nothing is held constant across variants, so the conversion data is noise. The team learns nothing from the spend.

The second failure is beautiful AI. Optimize for production polish (cinematic lighting, photorealistic faces, slick transitions) instead of for the curiosity gap. The output looks like a brand campaign. CTRs are flat because nothing about it stops a thumb. The team confuses aesthetic quality with performance quality.

Both failure modes share the same root cause. The team treats AI as a strategy tool instead of a production tool, and pulls humans out of the loop at exactly the points where humans matter most. The rest of this guide is what humans do, chapter by chapter, to keep that from happening.

Free download: the AI-Driven Creatives Skill Pack

Every prompt in this guide, plus four agent-like skills (hook-maker, script-maker, visual-director, creative-curator), packaged for one-click install into Claude Code or copy-paste into any LLM. The 3C Principle, the 4-Layer Hook System, BANP, the eight iteration axes, and the three-phase curation checklist, all in one downloadable pack.

Get the AI-Driven Creatives Skill Pack on GitHub

Chapter 1: Briefing AI Hooks

Most performance teams treat AI as a hook generator. That is the wrong job to give it. AI is a hook variant machine, and the brief you hand it is where 80% of the performance is decided. The principles of what makes a hook work have not changed in fifteen years of direct response, and they will not change because a model got cheaper. What has changed is throughput. The teams winning at AI-driven creative are not briefing looser, they are briefing tighter. They write prompts the way they used to write briefs to a senior copywriter, with constraints, framework names, and negative space. The teams losing are typing “give me 10 hooks for my fitness app” and wondering why every variant sounds like the same LinkedIn post.

This chapter is about the framing you bring to the model. Get the brief right and you can generate 200 hooks in an afternoon, half of which are testable. Get it wrong and you generate the same hook 200 times.

The three hook layers

When marketers say “the hook,” they almost always mean one of three different things. Treating them as the same thing is the most common drafting failure on AI-generated creative.

A mockup of a five-star App Store review is a visual hook that happens to use text. It is not a text hook. A text overlay that says “I tried this for 30 days” is a text hook, not a spoken hook, and it should not be duplicated by the VO. When you brief AI, name the layer you want. “Generate text overlays” and “generate VO opening lines” are different prompts and should never be combined into one.

The 3C Principle

Every hook, regardless of layer, has to do three things in the first second or two. We call these the 3Cs. If a hook is missing any one of them, it fails.

• Context. Who is this for, what category does it sit in. The brain has to categorize the content instantly or it scrolls. A hook for a sleep tracking app needs to communicate “this is about sleep” inside the first beat, even if the literal word “sleep” never appears.
• Clarity. What is this video about, why is it worth watching the next two seconds. Precision increases retention. Vague hooks (“you won’t believe what happened”) test worse than specific ones every time.
• Curiosity. An open loop. Tension, a contrarian angle, an implied mistake, an unresolved outcome. Without an open loop the viewer has no reason to keep watching, even if the first two Cs are perfect.

When you brief AI, you have to demand all three, by name. Models will happily generate hooks with curiosity and no clarity, or clarity and no context. Telling the model “must satisfy all three of Context, Clarity, Curiosity” inside the prompt cuts your reject rate roughly in half on the first pass.

The 4-Layer Hook System

Strong hooks operate across four layers at once. The strongest creative we ship stacks them intentionally, not accidentally.

A meditation app hook with a calm voice-over, a gentle riser, and a high-contrast text overlay is layered. The same VO over a static talking head with no overlay is a one-layer hook, and it will lose to the layered version even if the script is identical. When prompting AI, brief one layer at a time, then have a human stack them. Models are bad at multi-layer coherence and good at single-layer variants.

Text hook rules

These rules apply whether the human or the AI is drafting. They are non-negotiable inputs to the prompt.

1. First-person POV. Write as the user, not about the user. “I tracked my sleep for 30 nights and one number changed everything” beats “Users who track their sleep see results.”
2. Conversational, not staccato. “No notes, no typing, I just listen” reads like a person. “No notes. No typing. I just listen.” reads like a robot. AI defaults to fragments. Forbid them in the prompt.
3. Curiosity gap. Withhold the answer. The hook is the setup, not the reveal. “There’s a reason I stopped using my budgeting app at 11pm” beats “Budgeting at night made me overspend.”
4. Specific numbers and stakes. “I lost 14 pounds and gained back 9” outperforms “I lost weight and gained some back.” Specificity reads as true.
5. Length cap. Text overlays under 15 words. Video text overlays in the 14 to 20 word range maximum, readable in 1 to 2 seconds. Anything longer is an infographic, not a hook.
6. Natural speech texture. Contractions, fillers cut, but rhythm intact. “I’d been on three dating apps for a year” not “I had been using three dating applications for one year.”
7. Pair with the spoken hook, do not duplicate it. If the overlay says it, the VO should not repeat it. The two layers should triangulate the same idea from different angles.
8. Setup, never reveal. The hook opens the loop. The body closes it. A hook that gives the answer (“This app helped me sleep better”) has nothing to keep the viewer watching.

Hook patterns that work for direct response

These are the patterns we keep coming back to. Use them as scaffolding, not as fill-in-the-blank Mad Libs. Adapt the surface, keep the structure.

Sample copy-paste prompts for AI hook generation

These are the prompts. Substitute the placeholders, paste into Claude or ChatGPT, and you get hooks that need light editing instead of full rewrites.

Prompt 1: Text overlays for video ads

You are writing first-frame text overlays for a paid social video ad.

Product: [YOUR APP], a subscription app for [YOUR CATEGORY, e.g. sleep tracking].
User problem: [YOUR USER PROBLEM, e.g. people don't trust their sleep score and don't know what to change].

Generate 10 text overlay variants that follow these rules, no exceptions:
- First-person POV. Write as the user, not about the user.
- Curiosity gap. The overlay is the setup, never the reveal. Withhold the key fact (the what, why, or how).
- Under 15 words each. Must be readable in 1 to 2 seconds.
- No fragment stacking. No "One tap. That's it. No phone." rhythm.
- No product name in the overlay.
- No generic CTAs ("Download now", "Try free").
- Each must satisfy the 3C Principle: Context (category is recognizable), Clarity (worth watching the next 2 seconds), Curiosity (open loop).

Use a mix of these patterns, no more than 2 of any one pattern:
incomplete revelation, contrast/contradiction, teased outcome, "POV:" framing, direct challenge, question.

Output as a numbered list. After each overlay, in parentheses, name the pattern used.

Prompt 2: Verbal/spoken opening lines (first VO line)

You are writing the first spoken line of a paid social video ad. This is the verbal hook, not the on-screen text.

Product: [YOUR APP], a subscription app for [YOUR CATEGORY].
User problem: [YOUR USER PROBLEM].
Speaker persona: [e.g. 32-year-old woman, casual, talking to camera].

Generate 8 opening VO lines that follow these rules:
- First-person, conversational, sounds like talking to a friend.
- Flowing sentences, not staccato fragments. "No notes, no typing, I just listen" beats "No notes. No typing. I just listen."
- Length: 6 to 14 words. Must be deliverable in roughly 2 to 3 seconds at natural pace.
- Curiosity gap inside the first sentence. The viewer should think "wait, what?" before the second line.
- No "Hey guys," no "What's up," no filler openers.
- Do not name the product in the opening line. The product enters later in the script.
- Structure target: Audience + Problem or Desire + Unexpected Angle + Implied Outcome.

Output as a numbered list. After each line, in parentheses, note which element is doing the curiosity work (contradiction, withheld answer, surprising stat, etc.).

Prompt 3: Hooks for static display ads (single-frame)

You are writing hero text for a static (single-image) display ad on Meta or TikTok.

Product: [YOUR APP], a subscription app for [YOUR CATEGORY].
User problem: [YOUR USER PROBLEM].
Visual context: the ad will pair this text with [DESCRIBE PLANNED VISUAL, e.g. a phone screen showing a chart, a destruction-theme image, a mockup of a 5-star review].

Generate 7 hero text variants that follow these rules:
- Text must work without sound, without sequence, and without a payoff frame. The static is the entire ad.
- First-person POV where natural, second-person ("you") allowed for direct challenges.
- Under 14 words. Must be legible at thumb-stop distance on a phone.
- The text should pair with the visual to create a single curiosity gap. If the visual shows a result, the text raises the question. If the visual asks the question, the text complicates it.
- Do not name the product. The logo/CTA does that job at the bottom of the creative.
- No stacked fragments. No "infographic copy" (e.g. "$2,015 → 0 results").
- Must have all three of Context, Clarity, Curiosity.

Output as a numbered list. After each line, note in parentheses what the paired visual would need to do for this text to work.

Prompt 4: “Credential + Contradiction” pattern hooks

You are writing hooks in one specific direct-response pattern: "[Credential or identity]. [Contradiction]."

The structure: a one-line credential or identity statement, followed by a one-line contradiction that should not logically be true given the credential. The contradiction creates the curiosity gap.

Examples of the pattern (from other categories, do not reuse):
- "Certified personal trainer. Hated every workout app I tried."
- "I'm the worst writer in my office. My boss thinks I take the best notes."

Product: [YOUR APP], a subscription app for [YOUR CATEGORY].
User problem: [YOUR USER PROBLEM].
Possible user identities/credentials: [LIST 3 TO 5, e.g. "20-year meditator", "ICU nurse", "former insomniac", "marathon runner"].

Generate 10 hooks in this exact pattern, drawing from those identities. Rules:
- Two short sentences, total under 18 words.
- The credential must be specific and high-authority for the category.
- The contradiction must be a real, surprising tension, not a manufactured one. "Marathon runner. Can't sleep through the night." works. "Marathon runner. Likes our app." does not.
- First person, conversational.
- No product name.
- The contradiction is the setup, the body of the ad is the reveal. Do not give the answer.

Output as a numbered list. For each, note in parentheses which identity it uses and what the implied gap is (i.e. why the contradiction creates curiosity).

What NOT to do (text hook anti-patterns)

The reject pile from any AI hook generation pass is full of the same six failures. Train yourself to spot them on the first read.

• Fragment stacking. “One app. One tap. That’s it.” Reads like a robot. Rewrite as one flowing sentence.
• Statement that gives the answer. “This app helped me lose 14 pounds.” Loop is closed, no reason to watch. Rewrite as setup: “I lost 14 pounds and the thing that worked wasn’t the diet.”
• General or third-person. “Users who track their sleep see better results.” Nobody talks like this. Move to first person.
• Vague. “You won’t believe what happened when I tried this.” No context, no clarity. The viewer cannot categorize what they are watching.
• Over-engineered, infographic-style. “$2,015 → 0 results captured” reads like a Powerpoint slide, not a hook. Hooks are sentences, not data viz.
• Product-named in the hook. “I tried [App Name] for 30 days.” The product enters later. Naming it in the hook collapses the curiosity gap before it opens.

Chapter 2: Visual Elements at AI Scale

Most teams think AI image and video tools mean they finally get to skip the “what visual works” question. They don’t. AI doesn’t change which visual elements work on a paid feed. It changes how cheaply you can produce them, and that means the cost of picking the wrong format is now higher, not lower. If you brief a model badly, you don’t get one bad ad, you get forty bad ads in an afternoon. The teams pulling ahead right now are doing the opposite of what you’d expect: they spend more time upfront on format choice and visual hook taxonomy, then use AI to flood the variant slot. Pick the format for the message first, brief the AI second, render at volume third.

The 5 visual hook techniques

The visual hook is the first subconscious interrupt in a feed. It does not need to be cinematic. It needs to be noticeable. There are five techniques that consistently produce stop-the-scroll output, and each one has a different relationship with AI generation.

1. Dynamic Movement. Formula: Static Shot + Sudden Zoom (0.3 to 0.8s) = Immediate Pattern Break. A static talking head is low interruption. The same talking head with a quick zoom and a hand motion in the first second is significantly stronger. AI generation here is the easiest win of the five: image-to-video tools (Runway Gen-4, Sora 2, Pika, Kling) are explicitly built around camera moves. You can take a single still and generate a push-in, pull-back, or whip pan in under a minute. Brief the camera move directly in the prompt.

2. Color & Contrast. Formula: Base Visual + 15 to 25% More Contrast + Slight Saturation Boost = Increased Feed Standout. Feeds are visually similar by default. You stand out by punching saturation, increasing brightness, or planting one bold color element in frame. AI is good at this if you ask. Most marketers don’t ask. Add “high contrast, saturated, single bold color element (red phone, yellow sticky note, neon green dollar sign) against muted background” to every prompt and your hit rate climbs.

3. Context-Rich Objects. Formula: High-Meaning Object + Close Framing + Movement = High-Clarity Visual Hook. Objects that carry built-in narrative weight: cards, statements, phones showing dashboards, receipts, premium items, things that signal money or status. AI image models render these well, but they hallucinate text on screens, garble logos, and warp small UI elements. Plan for object-only framing (the back of the phone, the edge of the card) or composite the screen content in post.

4. Unexpected Imagery. Visual confusion that makes viewers think “what is happening” and stay. AI is the cheat code for this technique. The whole point of generative models is that they produce imagery that doesn’t exist in stock libraries. Lean in. Surreal scale, impossible compositions, an everyday object in the wrong context: cheap to produce, high to interrupt.

5. The Visual Hook Multiplier. Stack the previous four. Zoom-In + High Contrast + Curiosity Object + Audio Cue = High Interruption Probability. Brief the AI to produce all four at once: a saturated, close-cropped shot of a context-rich object with a built-in camera move. Then layer audio (notification ping, riser, hard cut to silence) in post.

Six static format conventions that work

These are the six static and short-form video formats that consistently work for subscription apps. AI is a force multiplier for all six, but each has a different production pipeline.

The deeper principle: most apps get trapped in one format because it’s winning. All video, no static. All product mockup, no UGC. Breaking format unlocks new audience segments, even when the underlying message stays the same.

Visual elements you can layer in

Once you’ve picked the format, these are the layers that compound on top. Each one used to be a production line item. Most are now near-free.

• Text overlays. Still the highest-leverage layer. AI doesn’t help much in production (Figma, CapCut, native editors are faster), but Claude or GPT can generate twenty overlay variants from a single body script in seconds.
• Stickers and emojis. Native-feeling visual texture. Free. Use them to break up a wall of text or to punctuate a transition.
• GIFs. Pre-AI workhorse, still works as a reaction beat in the middle of a video. Giphy is fine. No AI needed.
• Animations and AR. AI image models now produce frame-coherent short loops via Runway, Sora, and Kling. AR effects (Spark, Lens Studio) still require manual setup, but the assets that go inside them are AI-generated.
• Green-screen scene swap. Used to be a half-day editing job. Runway and Sora do background replacement in minutes. Brief the new background as its own prompt, composite, done.
• TTS. ElevenLabs voices are now indistinguishable from human reads at a conversational pace, and they cost cents per render. Use TTS for variant testing where you want to swap one phrase across thirty videos without re-recording.

Sample copy-paste prompts for AI visual generation

These are production prompts. Paste them, render, ship. Adjust the bracketed variables for your vertical.

(a) Discovery-format static (Midjourney or Gemini 3 Pro Image)

A cinematic close-up photograph of a young woman's hands wrapped around a ceramic coffee mug at a sunlit kitchen table, soft morning light streaming through a window, shallow depth of field, the mug fogged with steam. Her face is intentionally out of frame (cropped above the chin) so no facial features are visible. On the table next to the mug: a half-open paper journal with handwriting blurred, a pair of reading glasses, a small potted plant. Color palette: warm cream, soft beige, muted terracotta, with a single bold mustard-yellow napkin folded on the right side of the frame for color punch. Style: modern editorial lifestyle photography, Kinfolk magazine aesthetic, natural film grain, 35mm look. Aspect ratio 4:5 (Instagram and Meta feed). High resolution, sharp focus on the mug, soft bokeh on background. Leave the upper third of the frame as clean negative space (slightly out-of-focus wall) for a text overlay to be added in post. No faces, no logos, no text in the image, no people fully visible, no phones or screens. Quality: photoreal, magazine-grade.

(b) Transformation-format short video (Runway Gen-4 or Sora 2, image-to-video)

Image-to-video transition. Start state: an overhead shot of a cluttered home desk, papers scattered, three coffee cups, a closed laptop covered in sticky notes, dim afternoon light, desaturated cool tones (muted blues and grays), a sense of overwhelm. End state: same desk, same angle, but cleared and organized, one open notebook with neat handwriting, a single mug of fresh coffee, a small green plant added, warm golden-hour light, saturated warm tones (cream, gold, soft green). Transition: smooth match-cut morph over 4 seconds, with subtle camera push-in toward the notebook in the final second. Total duration: 6 seconds. Pacing: contemplative, not frantic. Music style: minimal piano with a soft uplift at the 4-second mark, no vocals, royalty-free indie folk feel. Aspect ratio 9:16 (vertical for Reels and TikTok). No people in frame, no faces, no on-screen text (text added in post). High realism, cinematic color grade.

(c) Context-rich object visual (destruction theme, fictional finance app, Midjourney or Gemini 3 Pro Image)

A cinematic still-life photograph of a ceramic pink piggy bank smashed in half on a dark walnut wood floor, coins (pennies, nickels, dimes) scattered around it, one coin still spinning on its edge in the foreground (motion blur on that single coin only). Side lighting from a window-left, deep shadows, dramatic chiaroscuro. The piggy bank's broken edges are clean and sharp, ceramic shards visible. A single crumpled receipt lies next to the pieces, slightly out of focus. Color palette: deep walnut brown, dusty pink, copper, with one bright element (a folded twenty-dollar bill peeking from under the wreckage) for visual punch. Style: still-life editorial photography, dramatic and slightly somber, like a Wall Street Journal feature image. Aspect ratio 1:1 (square). Sharp foreground focus, soft background bokeh. Leave the top quarter of the frame as clean dark negative space for a text overlay. No faces, no hands, no people, no logos, no text in the image, no brand marks on the coins or bill. Photoreal, high detail on ceramic texture and coin metal.

(d) Wall-of-Text static (Figma + AI text generation, with optional Midjourney background)

Build in Figma. Canvas: 1080 x 1350 px (4:5 Meta feed).
Background: a slightly off-white paper texture (download from Midjourney with the prompt: "lined yellow legal pad page, top-down flat-lay photograph, slightly worn edges, soft natural shadow, no text, no writing, ultra-realistic paper grain, 1080x1350, no people, no hands"). Place this image as the full background.
Foreground text: handwritten-style font (use Caveat, Kalam, or Patrick Hand from Google Fonts) at approximately 48pt, dark navy ink color (#1a2540), left-aligned with a 120px margin from the left edge.
Text content (generate with Claude or GPT using this prompt: "Write a 6-line handwritten-style note from a [vertical, e.g., language-learning app] user listing things they stopped doing once the app became a habit. Each line under 8 words. Conversational, first-person, slightly self-aware tone. No product name. No em dashes. Number each line."):
   1. Stopped pretending I'd "start tomorrow"
   2. Stopped buying textbooks I'd never open
   3. Stopped feeling guilty on my commute
   4. Stopped lying about my level
   5. Stopped translating in my head
   6. Started actually thinking in it
Add a small subtle doodle (an arrow, a star, an underline) next to line 6 in the same ink color. Add one realistic coffee-cup stain ring in the bottom-right corner (5% opacity). Export as PNG. No faces, no people, no logos, no product name in the text.

Format portability: the at-scale principle

Once a concept wins in one format, you don’t reinvent it. You remix it. Same message, different format, different audience segment. A Discovery-format video that wins on TikTok becomes a Wall-of-text static for Meta, a Listicle carousel for LinkedIn, a Split-screen for YouTube Shorts. The thesis stays constant. The container changes. AI makes this remix loop close in days instead of weeks, because the prompt for each new format is a one-line tweak away from the prompt that made the original. The teams that win don’t have more ideas, they have more containers per idea.

Visual hook checklist (pre-publish)

Before approving any creative, run through these six questions. If three or more are weak, rewrite the hook.

• [ ] Does something move in the first second?
• [ ] Is there visual contrast or color strength?
• [ ] Is context clear within 1 to 2 seconds?
• [ ] Is there a defined audience?
• [ ] Is there an open loop?
• [ ] Is text readable and native?
• [ ] Is audio clean and supportive?

Chapter 3: Generating Scripts That Perform

Most teams using AI to write ad scripts get sloppy output, and they blame the model. The model isn’t the problem. The framework is. When you prompt Claude or ChatGPT for “a Facebook ad script,” it reaches for what it saw most often in training: PAS, BAB, AIDA, the same three or four frameworks every copywriting blog has been recycling for two decades. Generic input, generic output. The win is not a better model. The win is constraining the model to a tighter, performance-tested structure your competitors aren’t using. We use one called Before / After / Now / Punchline (BANP). It’s the spine of every script we ship, and it’s the single biggest unlock when you put AI in the loop.

The Before / After / Now / Punchline structure

Four parts, in order:

1. Before (the problem/status quo): What life looked like before the solution.
2. After (the shift): What changed once the user found the product.
3. Now (the current state): The specific behavior or result today.
4. Punchline (the payoff): A memorable closer that reinforces the value.

Here’s how it reads for a productivity app that records and transcribes meetings:

• Before: “I used to spend 20 minutes after every meeting trying to remember what was said.”
• After: “Now I just hit record on my phone and slip it in my pocket.”
• Now: “It transcribes everything and gives me a perfect summary.”
• Punchline: “My coworkers think I have a photographic memory. I don’t correct them.”

Same structure for a budgeting app:

• Before: “I’d open my banking app on Sunday night and feel my chest tighten.”
• After: “Then I started letting one app pull every transaction in automatically.”
• Now: “Every Sunday I get a one-screen view of where my money actually went.”
• Punchline: “I haven’t dreaded a Sunday night in six months.”

And once more, for a meditation app:

• Before: “I’d lie in bed at 1 a.m. doing breathing exercises I’d read about on Reddit.”
• After: “A friend told me to stop guessing and just press play on something.”
• Now: “I queue up a 12-minute session and I’m out before it ends.”
• Punchline: “I have not finished a single one. That’s the point.”

The structure is simple. The discipline is in not skipping a beat.

Body copy rules that don’t change with AI

These rules predate AI and survive it.

• Length: 20 to 25 seconds when read aloud. Typically 45 to 65 words. Never over 70.
• Tone: Conversational, first-person. Sounds like someone telling a friend, not reading a script. No corporate-speak, no filler.
• Flow: Flowing sentences, not choppy two and three word fragments.

The read-aloud test (mandatory). Before any script ships, read it out loud at natural pace. If you stumble, rewrite. Four things to catch:

• Dangling phrases. “Every action item, pulled out.” reads incomplete. “Every action item” is cleaner.
• Fragments that don’t connect. “One app. One tap.” reads choppy. “One app, one tap” flows.
• Missing context. “5 minutes after” leaves you asking after what. “Five minutes later” is complete.
• Stacked short fragments. “No notes. No typing. I just listen.” reads like a robot. “No notes, no typing, I just listen” reads like a person.

What to avoid. AI-sounding fragments (“One tap. That’s it. No phone. No laptop.”), generic claims without specificity, product feature lists disguised as stories, and sentences that start with the product name. If your script opens with “[App Name] helps you…”, scrap it.

Hooks for video scripts (the verbal hook)

The verbal hook is the first VO line. It does the same job a headline does in display: stops the scroll and forces a “wait, what?” The structure:

Audience + Problem/Desire + Unexpected Angle + Implied Outcome

Examples across verticals:

• “If you’re running Meta ads and your CPA keeps rising, this creative mistake is probably why.”
• “If you keep starting language apps and quitting in week two, the streak isn’t the problem.”
• “If you’ve trained five days a week for a year and still don’t see your abs, your workout isn’t broken. Your kitchen is.”
• “If you check your bank balance before every coffee, you don’t have a spending problem. You have a visibility problem.”

Patterns that work for spoken delivery, adapted from our hook table:

Delivery rules: start immediately, no “Hey guys,” confident pacing, no filler. Record five to ten variations of any hook. Hooks are variables, not finals.

Sample copy-paste prompts for AI script generation

These are the centerpiece. Paste into Claude or ChatGPT, fill the bracketed placeholders, ship.

Prompt A: Generate 3 BANP scripts from one concept brief

You are a direct response copywriter for a subscription app called [YOUR APP].
The user problem you solve: [YOUR USER PROBLEM]
The angle for this concept: [ANGLE / EMOTIONAL HOOK]
The target audience: [AUDIENCE DESCRIPTION]

Write 3 video ad script variants using this exact four-part structure, in order:
1. Before: what life looked like before the user found the product
2. After: the shift, what changed when they found it
3. Now: the specific behavior or result today
4. Punchline: a memorable closer that reinforces the value

Constraints:
- 45 to 65 words per script. Never exceed 70.
- Conversational, first-person. Sounds like someone telling a friend.
- Flowing sentences. No stacked two and three word fragments.
- No sentence starts with the product name.
- No generic claims ("save time", "feel better"). Be specific.
- No feature lists. Tell a story.
- Do not use PAS, BAB, AIDA, or any other framework. Only Before / After / Now / Punchline.
- Do not use em dashes or en dashes.
- Each script must pass a read-aloud test: no dangling phrases, no broken fragments, no missing context.

Output the 3 variants numbered, with each part labeled (Before / After / Now / Punchline).

Prompt B: Generate 5 verbal hooks for one concept

You are writing the opening VO line (verbal hook) for a video ad for [YOUR APP].
Concept: [CONCEPT NAME, ONE SENTENCE]
Audience: [AUDIENCE]
Core tension or surprising angle: [ANGLE]

Write 5 verbal hook options, each 1 sentence, each under 18 words. Use this structure:
Audience + Problem/Desire + Unexpected Angle + Implied Outcome

Vary the 5 across these patterns (one per hook):
1. "If you've ever [relatable frustration], [implied solution]"
2. "I [credential/identity]. [contradiction]."
3. "I'll show you how to [surprising action]"
4. "When your [authority figure] [does frustrating thing]..."
5. "I [did impossible thing] and [surprising result]"

Constraints:
- No "Hey guys", no "Did you know", no "What if I told you".
- Do not name the product in the hook.
- Each hook must open a curiosity gap. The viewer should think "wait, what?" or "how?"
- No em dashes or en dashes.
- Read aloud test: each must sound natural in 3 seconds.

Output 5 numbered hooks, labeled with the pattern used.

Prompt C: Rewrite a flat AI-generated script

Below is a draft ad script that reads like generic AI output. Rewrite it.

Apply this exact four-part structure: Before (problem before product), After (the shift), Now (current behavior or result), Punchline (memorable closer).

Fix all of the following:
- Choppy two and three word fragments. Replace with flowing sentences.
- Dangling phrases that read incomplete.
- Stacked short fragments that read like a robot.
- Missing context (e.g., "5 minutes after" with no anchor).
- Sentences that start with the product name.
- Generic claims with no specificity.
- Feature list framing instead of story framing.

Constraints:
- 45 to 65 words. Never over 70.
- Conversational, first-person.
- Do not use PAS, BAB, AIDA. Only Before / After / Now / Punchline.
- No em dashes or en dashes.
- Must pass a read-aloud test at natural pace.

Original draft:
[PASTE FLAT SCRIPT HERE]

Output the rewritten script, with each part labeled, plus a one-line note on what you changed and why.

Prompt D: 3 audience segment variants of the same concept

You are writing 3 ad script variants for [YOUR APP] using the same product and angle but 3 different audience segments. The audiences are:

1. New user: [DESCRIBE NEW-USER STATE]
2. Switcher (coming from a competing approach or app): [DESCRIBE SWITCHER STATE]
3. Power user: [DESCRIBE POWER-USER STATE]

For each audience, write one script using this exact structure:
- Before (specific to that segment's pre-product state)
- After (the shift, specific to that segment)
- Now (the current behavior, specific to that segment)
- Punchline (segment-relevant payoff)

The After / Now / Punchline can be similar across segments. The Before must be sharply different. That's the point of the exercise.

Constraints:
- 45 to 65 words each.
- Conversational first-person.
- No sentence starts with the product name.
- No PAS, BAB, AIDA. Only Before / After / Now / Punchline.
- No em dashes or en dashes.
- Read-aloud tested.

Output 3 scripts, labeled by segment.

Three common mistakes teams make with AI scripts

1. Asking the AI to write the script before locking the angle and Before state. AI is not a strategist. If you hand it a vague brief (“write 3 ads for our fitness app”), you get vague output. The angle, the audience, and the specific Before state are human decisions. Lock them, then prompt. The quality ceiling of the output is set the moment you write the brief.

2. Accepting the first output instead of running the read-aloud test. AI scripts read fine on the page. They fall apart out loud. Stacked fragments, dangling phrases, robotic transitions, the model ships them constantly because it’s optimizing for plausible written English, not natural spoken English. Read every variant aloud. Cut what stumbles. This single step lifts script quality more than any prompt tweak.

3. Letting the AI default to its training-data favorite. Without an explicit constraint, the model reaches for PAS, AIDA, or a generic feature list. It will do this even if you ask for “something different.” The fix is to name BANP in the prompt, define each part, and explicitly negate the alternatives. Tell the model what not to do. Generic frameworks are the gravity well, and you have to push out of it every prompt.

What this changes vs. what stays the same

AI accelerates production. One person can now generate fifteen script variants in the time it used to take to draft three. That’s real. What hasn’t changed: the structure that makes scripts perform, the angle that makes a concept worth shooting, and the human judgment that picks the Before state. AI fills the rest. Curation is mandatory. The teams winning with AI aren’t the ones generating the most variants. They’re the ones with the tightest filters on what gets shipped.

Chapter 4: Iterating Across the Eight Axes

This is the chapter that gets ten times bigger with AI. The UGC playbook listed eight axes you could iterate on for any winning concept. UGC let you cover three or four of them per testing cycle. AI lets you cover all eight, sometimes in a single afternoon.

The eight axes

1. Messaging. Text overlay variants, hook copy alternatives, comment-style hooks, question hooks. Same script, different framing on the first frame.
2. Background colors. Overlay color, video background tone, app-screen vignette, brand-accent color choice. Often underestimated. A change from cool blue to warm cream on the same composition can shift CTR materially.
3. Duration. 10s, 15s, 20s, 30s for video. Aspect ratio swap for static (1:1, 4:5, 9:16). Different durations attract different placements.
4. Music. Uplifting, suspense, ASMR-style, no-music. The track is the emotional spine. A meditative track on a bombastic script feels broken; an urgent track on a calm script feels fake.
5. Voice over. TTS, AI-cloned voice, original human, no VO at all. Each option tests differently with different audience segments.
6. Animation. Transition style, sticker entrance, text emphasis pattern, motion type (cut vs. fade vs. slide).
7. Scene order. Which beat opens the script, which closes it. Reordering Before/After/Now/Punchline as Now/Before/After/Punchline gives a different curiosity gap entirely.
8. Visual hook treatment. Theme variant, prop substitution, format swap (Discovery to Wall-of-Text on the same script).

How a concept fans out

Once a concept wins, the variant tree can run into the hundreds. Five hook copy variants, three visual treatments, two format choices, two music beds, two VO options, two scene orders. Multiply through and a single concept supports 240+ permutations.

You don’t ship all 240. You ship a strategically sampled subset that covers the dimensions you want to test, while holding everything else constant. Fifteen to twenty variants in active rotation per concept is usually the right ceiling. Beyond that, you’re diluting learning, not gathering it.

Test design for AI variants

For a concept entering test, run three cohorts in parallel:

1. Hook copy cohort. Same visual, same format, same script. Only the text overlay or first VO line varies. Tests which hook framing wins.
2. Visual hook cohort. Same hook copy, same format, same script. Only the visual treatment varies. Tests which theme wins.
3. Format cohort. Same hook copy, same visual treatment. Different format (Discovery vs. Transformation vs. Wall of Text). Tests which container wins.

Hold the strategy layer constant within a cohort. Vary the production layer only. That’s the discipline that makes the data readable.

After 5 to 7 days of spend at meaningful conversion volume, the winners across the three cohorts combine into the next iteration’s brief. Iterate again. Iterate again. Within four to five cycles, a strong concept compounds into a top performer.

The teams that don’t run this discipline end up running 80-variant tests where every variable shifts at once. The data tells them nothing, even though the spend looks “rigorous.” Cohort design is what separates iteration from churn.

Chapter 5: Safe Zones

This chapter is mostly unchanged from UGC. AI doesn’t change platform overlay rules. The like, comment, share, and caption areas on TikTok, Instagram Reels, Instagram Stories, and YouTube Shorts have specific zones the headline must not overlap. If your AI-generated overlay sits where the like button is rendered, it gets cropped on screen. The video looks broken.

Fix the overlay zone at brief time. Use 1080×1920 as the master canvas. Position critical text at least 250 pixels from the bottom edge for Instagram Reels, 340 pixels from the bottom for TikTok. Always preview each variant in the platform’s native preview tool before pushing live. The free safe-zone templates from the original UGC playbook still work; the AI-driven workflow uses the same templates.

Chapter 6: The AI Production Workflow

The chapter the UGC playbook spent ten pages on, planning and directing a remote shoot, gets replaced by something equally rigorous, just structurally different. There is no casting call. There is no shoot day. There are no creators on a Zoom waiting for direction. Instead there is a tool stack, a prompt library, an edit-versus-regenerate decision tree, and a curation step. The shape of the work is different, but the discipline required is the same. If anything, it is harder, because the cost of a bad iteration drops to almost zero, and that is exactly what makes most teams sloppy. When generation is cheap, curation becomes the only thing that matters. This chapter is the operator’s manual for that loop: which tool to open for which job, how to prompt it so the output is shippable instead of slop, when to fight a bad seed with edits, when to start over, and how the same concept fans out into twenty variants in an afternoon. Read this once with a notebook open. The five sample prompts at the end are meant to be pasted, not admired.

The shape of an AI-driven production cycle

Here is the loop, end to end, in five stages. Each stage has a job and a decision point.

BRIEF (human)  ->  GENERATION (AI)  ->  CURATION (human)  ->  EDIT (AI-assisted)  ->  SHIP

1. Brief (human). A one-paragraph concept that names the angle, the hook layer, the format, and the negative space (what the visual must NOT show). Decision point: is this brief specific enough that two different operators would generate the same kind of output? If not, rewrite the brief.

2. Generation (AI). Image, video, voiceover, music, copy, in whichever order makes sense for the format. Decision point: did the model understand the brief, or did it fill in defaults you did not ask for? If defaults leaked in, the brief was vague.

3. Curation (human). Sort the outputs into ship, edit, or trash piles. Decision point: would I run this against my best historical creative? If no, kill it. Most teams skip this step and ship everything that rendered.

4. Edit (AI-assisted). Inpainting, color correction, copy fixes, audio leveling. Decision point: can this be fixed with a five-minute edit, or am I about to spend two hours trying to rescue a fundamentally broken seed?

5. Ship. Push to ad account, label with concept ID, batch ID, and variant ID so you can read the report later.

The whole loop should run in hours, not days. If it is taking days, you are stuck inside one of the two human stages and it is not the AI’s fault.

Tools by use case (as of 2026)

The tool stack changes every quarter. The categories do not. Here is what to open for what job.

Images and stills. Visual hooks, app mockup hero shots, static format compositions, character generation if you absolutely need a person.

• Midjourney v7. Best for stylized, painterly, brand-feel work. Color palettes feel intentional. Composition leans cinematic by default. Fails at: precise text rendering inside the image, exact product replication, photoreal faces with eye contact. Right call when: you want a destruction-theme hook for a finance app, or a mood-piece visual for a meditation brand. Style first, accuracy second.
• Gemini 3 Pro Image. Best for fast, near-photorealistic scene generation with strong prompt adherence. If you describe a cluttered desk with seven specific objects, it gets six of them right. Fails at: stylization, painterly looks, anything that needs an art direction beyond photo-real. Right call when: you need a context-rich product-in-the-world shot for a productivity app, or a top-down flatlay for a fitness vertical.
• Sora 2 (still mode) and DALL-E equivalents. Best for image-to-image edits, inpainting, and scene extension. You generated a great hero shot but the overlay zone has a distracting object: this is the tool. Fails at: generating from scratch with the same finesse as Midjourney or Gemini. Right call when: you are 90% there and need to surgically fix one element.

Video. Motion visual hooks, scene transitions, b-roll clips, short generative segments to splice between app screen recordings.

• Sora 2. Short generative clips up to roughly ten seconds, with strong scene composition and physics. Fails at: long takes, consistent character across multiple shots, exact lip-sync for VO. Right call when: you need a six-second transformation visual or a scene transition that bridges two app screen recordings.
• Runway Gen-4 (latest gen). Best for image-to-video with motion control and camera moves. You have a static mockup and you want a slow push-in, a parallax slide, or a subtle camera dolly: this is the tool. Fails at: generating wildly novel scenes from text alone. Right call when: animating a phone mockup, animating a product still, adding subtle motion to a hero shot.
• Pika. Best for stylized animation, character motion, and looser, more illustrative motion work. Fails at: photorealistic motion, precise camera control. Right call when: you are doing an animated explainer-style hook for a language learning or education vertical.

Audio. Voiceover and music. The voice and the track decide whether the ad feels native or feels like an ad.

• ElevenLabs. Cloned voiceover, multilingual TTS, emotional register control (warm, urgent, conspiratorial, deadpan). Fails at: extreme emotional range like genuine crying or screaming, very long single takes without pacing drift. Right call when: you need a 25-second VO in a specific tone, or you need the same voice across ten variants for consistency.
• Suno. Stylized music tracks, including vocals, in named genres and moods. Fails at: rights clearance for specific paid placements (read the terms), and matching an exact reference track. Right call when: you want a custom track that does not sound like the same five royalty-free loops every other ad uses.
• Platform-native libraries (TikTok Commercial Music Library, Meta’s audio library). The safe, cleared-for-ad-use music option. Fails at: standing out, since every advertiser has access to the same catalog. Right call when: you need to ship today and you do not want to deal with rights questions.

Copy. Hooks, scripts, body copy variants, brief writing, negative-space construction.

• Claude Opus and Sonnet. Best for rule-following at length, body copy that respects guidelines, and brief writing where you want the model to honor a long list of constraints. Fails at: rapid-fire one-line ideation if you are not steering it. Right call when: writing a 25-second VO script that has to follow a Before/After/Now/Punchline structure, or generating ten body copy variants that all respect a do-not-say list.
• ChatGPT. Best for fast variant generation, brainstorming, and A/B prompt iteration. Fails at: holding a long, complex rule set across many turns. Right call when: you have a working hook and you want twenty quick rewrites in different tones.

The reader rule: open Midjourney for mood, Gemini for accuracy, Sora for motion, Runway for animating stills, ElevenLabs for voice, Claude for rules, ChatGPT for speed.

The negative space rule (the prompting principle that separates shippable from slop)

Most prompting advice talks about specificity. Specificity is half the answer. The other half, the half that separates shippable from slop, is what you forbid. AI rewards specificity AND it rewards constraints. Tell the model what NOT to do, in plain language, in the same prompt.

Examples of negative space that should appear in almost every prompt: no faces, no hands (unless hands are critical and you are willing to inpaint later), no actor looking shocked or surprised, no generic stock-photo backgrounds, no stock-music-sounding tracks, no invented statistics, no on-screen text the model will mangle, no logos other than the one I provide, no AI artifact tells like extra fingers or melted text.

The negative space is where the strategy lives. A meditation app’s brief that says “calm, peaceful, soft lighting” gets the same output every other meditation app gets. A brief that says “calm, peaceful, soft lighting, no spa imagery, no stock yoga poses, no white candles, no bare feet on hardwood, no lavender” forces the model to find a fresher visual. The forbidden list is the brief.

If your prompt does not have a “do NOT include” section, you are leaving the most strategic part of the work to the model’s defaults. The model’s defaults are everyone else’s defaults.

Sample copy-paste prompts for visual generation

Five prompts. Pasteable today. Substitute your vertical and your specifics.

(a) Midjourney prompt, destruction-theme visual hook

A premium fitness watch with a deeply cracked screen, lying face-up on a dark slate surface, faint shards of glass scattered around it, a single shaft of cold blue light cutting across the frame, cinematic lighting, shallow depth of field, hyper-detailed product photography, moody and editorial in the style of a high-end magazine still life, 4:5 aspect ratio, dark background, no faces, no hands, no human presence, no logos other than a generic blank watch face, no text overlays, no stock-photo gradients, no studio softbox look --ar 4:5 --style raw --v 7

(b) Gemini 3 Pro Image prompt, cluttered-desk pain-point shot

Top-down photograph of a chaotic home office desk at 11pm, lit by a single warm desk lamp from the upper left and the cool blue glow of a smartphone screen showing 47 unread notifications. On the desk: a half-drunk cup of coffee with a ring stain on a paper napkin, three crumpled sticky notes with handwriting, an open laptop with too many browser tabs, a stack of unopened bills, a tangled phone charger, a closed paper planner. Realistic textures, slight motion blur on the steam from the coffee, photo-real lighting, 1080x1920 vertical aspect ratio. Do NOT include: any human, any hands, any face, any logos, any readable brand names, any clean minimalist desk aesthetic, any plants, any stock-photo styling, any soft pastel palette.

(c) Sora 2 prompt, six-second transformation video

A six-second vertical 9:16 video. First three seconds: a tense woman seen only from behind, hunched at a cluttered desk in a dim apartment, shoulders tight, the warm-cool color clash of a desk lamp and a laptop screen on her face (face never visible), papers and a phone vibrating on the desk. Camera holds steady on a medium shot. At the three-second mark, a smooth match-cut to: the same woman from behind, now standing at a tall window in soft natural morning light, shoulders relaxed, holding a plain ceramic mug, looking out at out-of-focus trees. Camera pushes in slowly over the final three seconds. Pacing: hold-cut-push. Cinematic, naturalistic, color graded warm. Do NOT include: visible face, on-screen text, music suggestions, any meditation app UI, any other people, any stock-yoga imagery, any spa cliches, any candles or incense.

(d) Runway Gen-4 image-to-video prompt, animating a static phone mockup

Input: the attached static 9:16 phone mockup image. Animate as a five-second clip. The phone holds steady in frame for the first half second. Then the phone screen content slides smoothly from screen one to screen two to screen three, each screen visible for roughly 1.2 seconds, with a soft horizontal swipe transition between them. While the screens slide, the camera does a subtle slow push-in on the phone, roughly 8 percent zoom over the full duration. Keep the phone bezel and background plate completely static. Lighting should match the source image. Motion should feel intentional and product-demo clean, not floaty or AI-drifty. Do NOT add: any text overlays, any hands holding the phone, any background motion, any color shift, any reflections that were not in the source.

(e) ElevenLabs prompt, 25-second voice-cloned VO take for a sleep tracking app

Voice settings: cloned voice, emotional direction, “warm, conspiratorial, slightly amused, like a friend telling a secret over coffee, low and steady, never urgent.” Stability around 40, similarity around 75, style exaggeration low. Read the script below at a relaxed pace, with a clear breath before the punchline.

You know that thing where you wake up at 3 a.m. and your brain is suddenly running a TED talk about an email from last Tuesday? Yeah. For two years I thought I was just bad at sleeping. Turns out I was bad at knowing what was wrecking my sleep. The screen time, the late coffee, the room that was three degrees too warm. Once I could see the pattern, I could actually fix it. Took about a week. The app does not put you to sleep. It just shows you why you are not sleeping. Most people figure it out in their first three nights.

Edit vs. regenerate: the decision tree

Most teams over-regenerate. They get a render that is 90% right, the overlay zone has a slight conflict, and they hit generate again, hoping the next roll of the dice fixes it. It usually does not. It usually breaks something else that was working. Then they roll again. Three hours later they are further from shippable than they were at minute fifteen.

The discipline is the opposite. Edit until you cannot, then regenerate.

Regenerate if: the subject is wrong, the scene is wrong, the tone is wrong, the model fundamentally misunderstood the brief, the composition is unusable, or the style is off in a way that no edit can rescue. These are seed problems. A new seed is the only fix.

Edit if: 90% is right and one element is wrong. The overlay overlaps the like-zone on TikTok. The accent color is off-brand. The dollar value or stat in the on-screen text is wrong. A hand position is unnatural. A logo is in the wrong place. A piece of clutter is distracting. There is one extra finger. The color grade is half a stop too cool.

Tools for edits. Inpainting in Sora or DALL-E for surgical content swaps. Photoshop or Figma for color correction, layout, and overlay placement. After Effects or CapCut for video compositing, trims, speed ramps, and adding the platform-native UI cues that make the ad feel native. ElevenLabs for re-cutting a single line of VO without redoing the whole take.

The rule of thumb: if the fix is fifteen minutes of editing or less, edit. If the fix would take longer than regenerating, regenerate, but with a sharper brief than the one that produced the broken seed.

Common failure modes when prompting AI tools

1. Vague subject. The prompt says “a woman in a kitchen.” The fix: specify exact subject, scene, props, lighting, mood, time of day, camera angle, and lens feel.
2. Missing negative space. The prompt only says what to include. The fix: every prompt gets a “do NOT include” section. List the cliches, the artifacts, the off-brand elements.
3. Asking the model to do too many things in one prompt. Hero shot plus three product features plus on-screen text plus a logo. The fix: split into multiple prompts and composite the layers in Photoshop or After Effects.
4. Defaulting to “cinematic” or “photorealistic” without specifying style. The model picks an average. The fix: name a reference style, a film stock, a director, a photographer, a magazine editorial look. Specificity over genre.
5. Skipping the aspect ratio. The model defaults to square or 16:9. The fix: specify 4:5, 9:16, or 1:1 in every prompt, every time.
6. Letting AI generate faces when not needed. AI faces still fail eye contact in direct response placements, and the uncanny valley costs you the first second. The fix: brief faceless visual hooks. Hands, products, environments, over-the-shoulder shots, scenes without people.
7. Using AI footage where stock or app screen recordings would be cheaper and stronger. A real screen recording of your app beats a generated approximation every time. The fix: only use AI for what cannot be filmed or screen-recorded.
8. Not specifying duration or pacing for video. “A short clip” gets you a default. The fix: name the seconds, name the cuts, name where the camera moves and where it holds.

A worked example: one concept, three formats, twenty variants

Pick a sleep tracking app. The concept brief, written in one paragraph: “Angle: most people think they are bad at sleeping, but they are actually bad at knowing what is wrecking their sleep (screens, caffeine timing, room temperature). Hook layer: pattern interrupt with a relatable 3 a.m. moment, then a reveal that the cause is diagnosable. Format: vertical 9:16, 25 seconds, faceless visuals plus app screen recording, conspiratorial VO. Negative space: no spa imagery, no people in bed, no stock yoga, no generic moon-and-stars graphics, no claims about hours of sleep gained.”

Visual generation. Open Sora 2 for the cold-open transformation clip (six seconds, tense person at desk to calm person at window, the prompt from section c above). Open Gemini 3 Pro Image for the cluttered-desk pain-point hero shot (the prompt from b, adapted: a phone on the nightstand at 3 a.m. showing a sleep app notification, a glass of water, a half-read book, blue screen glow on the wall). Two source assets, ten minutes of generation time.

Script generation. Open Claude. Prompt: “Write a 25-second VO script for a sleep tracking app, structured as Before-After-Now-Punchline. Tone: warm, conspiratorial, slightly amused. Constraints: do not promise hours of sleep gained, do not name competitors, do not use the word ‘transform’, do not start with a question, do not use the phrase ‘game changer’. The hook must be a relatable 3 a.m. moment. The reveal must be that the cause is diagnosable, not that the user is broken.” Claude returns the e-prompt script above on the second pass. First pass started with a question, prompt got tightened.

Variant explosion. The same concept fans out across three formats and twenty variants. Format one: cold-open transformation video (six versions, swapping the opening visual, the VO emotional register, and the music bed). Format two: static-image-plus-VO, using the cluttered-nightstand Gemini render and four hook copy variants over the same VO. Format three: app-screen-recording-led, where the screen recording is the hero and the AI b-roll is the supporting layer (six versions, swapping which app feature is foregrounded). Twenty total variants, all from one concept brief, one Sora clip, one Gemini still, one Claude script, and one ElevenLabs voice.

Curation. Cut to ten percent. Walk through all twenty with a fresh eye, ideally the next morning, ideally with a second person on the team. Kill the ones that do not pass the “would I run this against my best historical creative” bar. Ship two per format, six total, into the ad account.

Total time. An afternoon. Not a week. Not a shoot day. One operator, one concept brief, four tools, six shipped variants, ready for the test framework in Chapter 8.

Workflow integration: where humans plug in

The brief is human. The curation is human. The strategic decisions, what to test, what to kill, what to scale, are human. The production is the AI’s job.

The humans-in-the-loop checklist (covered in detail in Chapter 7) plugs into this workflow at exactly two stages. It plugs in at the brief-writing stage, where a human pressure-tests the angle, the negative space, and the format choice before any tokens get spent. And it plugs in at the curation stage, where a human reads every output against the brand, the platform policy, and the brief itself, before anything ships.

Everything in between, the generation, the variant explosion, the inpainting, the audio leveling, the upscaling, is the AI’s job. The mistake most teams make is inverting this. They let the AI write the brief by defaulting to its training data, and they skip the curation because generation was so cheap it felt wasteful to throw outputs away. Both inversions kill performance. The brief is where strategy lives. The curation is where taste lives. AI cannot do either of those well, and the teams that win are the ones that protect those two stages from automation, while automating everything else aggressively.

Chapter 7: Humans in the Loop, the Drafting and Curation Checklist

The work AI eliminates is production work. The work AI does not eliminate is strategy work and curation work. In an AI-driven workflow, those two layers carry more weight than they did in UGC, because the iteration volume is so much higher.

This is the checklist we run for every AI-driven creative concept. Three phases: drafting (before any AI generation), production (during generation), and curation (after generation, before shipping).

Drafting (human, before any AI generation)

1. Lock the angle. Most subscription apps support three to five messaging pillars. For a fitness app: BodyTransformation, MentalDiscipline, ConvenienceWin, SocialAccountability, IdentityShift. Every concept maps to one. Without this lock, the AI generates angle-less variants.
2. Specify the hook layers separately. Text hook, visual hook, spoken hook. Each does different work. Conflating them is the most common drafting failure.
3. Write the text hook to spec. First-person POV. Conversational. Curiosity gap. Specific numbers and stakes. 14 to 20 words. Setup, not reveal.
4. Specify the visual hook as a theme, not a single image. Themes survive iteration; single images run out after one test.
5. Pick a fixed format up front. Discovery, Transformation, Progression, Split-screen, Listicle, or Wall of Text. Format is set at brief time, not generation time.
6. Write the body copy to spec. Before / After / Now / Punchline. 45 to 65 words. Conversational, first-person. Read-aloud tested.
7. List the negative space. What the AI is not allowed to do. No faces. No “actor looking shocked.” No invented stats. No stock-music-sounding tracks. No off-brand props or palette. The negative space is where the strategy lives.

Production (AI does the work, human supervises)

1. Generate on the eight axes simultaneously. Hook copy, background, duration, music, VO type, scene order, transition style, visual hook treatment.
2. Generate scripts using BANP, not training-data defaults. Explicitly forbid PAS, BAB, AIDA in the prompt. Constrain to Before / After / Now / Punchline.
3. Hold the strategy layer constant across variants. Vary the production layer only. This is what makes test results readable.
4. Stop generating when you have enough comparable variants. Twelve to twenty per concept across three formats is usually the right ceiling.

Curation (human, before shipping)

1. Read each variant aloud. If it sounds like ad copy, kill it. AI-generated copy fails this test more often than it passes.
2. Check the hook layers for redundancy. If the text hook and the spoken hook reveal the same fact, kill one.
3. Verify no fabricated specifics. AI invents stat percentages, customer counts, and dollar amounts when it has nothing else to say. Every number has to be defensible.
4. Check faces, if any. AI faces still fail eye contact in a way that reads uncanny in DR placements. Most of the AI-driven work that performs is faceless for this reason.
5. Check the safe zone. If the headline overlaps the like button, it gets cropped on screen.
6. Cut to ten percent. Most AI output is graveyard. Ship the best one to two variants per format per concept. Kill the rest before they touch an ad account.

The reason humans matter at all three phases is that the AI does not know the brand, does not know what’s failed in past tests, and does not know which DR principles are non-negotiable. It produces variants. Humans decide which variants represent the brand and which ones don’t.

Chapter 8: At-Scale Operational Discipline

Volume is the foundation. Twenty to forty creative variants per week is the threshold below which the three trends from earlier (audience diversity, format diversity, human-in-the-loop) can’t compound. AI makes the production side of that volume cheap. The operational discipline to make the volume useful is the part most teams miss.

Five operational habits separate the teams that scale AI-driven creative from the teams that ship a lot of bad ads.

1. Format portability. Once a concept wins in one format, adapt it to the others. Don’t reinvent. Remix. A Discovery-format video winning on TikTok becomes a Wall-of-Text static for Meta, a Listicle carousel for LinkedIn, a Split-screen for YouTube Shorts. Same message, different container, different audience segments engaged. The remix loop is what compounds learning across the format diversity trend.

2. Micro-audience targeting. Don’t ship one ad for “everyone who’d want this app.” Ship five concepts for five micro-audiences. For a productivity app: ESL professionals, parents of neurodivergent kids, freelancers, creative writers, ICU nurses. The named audience converts at materially higher rates than the abstract audience. AI lets you produce per-segment creative without per-segment shoot budget.

3. Naming conventions. Every variant gets a consistent name format encoding emotion, context, audience, batch, and concept. Without this, the data is unattributable and the test results are useless. Every file, every time. Even when it feels excessive. Especially when it feels excessive.

4. Themed store pages (CPPs). Ad creative does not exist in isolation. The landing page or App Store CPP must feel like a continuation of the ad. A Discovery-format ad targeting ESL learners should land on an App Store CPP themed for ESL learners. Break the chain between ad and landing surface, and CPI inflates by 30 to 50%.

5. Weekly batch cadence. Pick a day for concept planning (Monday) and a day for batch sync (Friday). Every week, the team plans what’s testing next, reviews what shipped, and reads the test results from the previous batch. The cadence locks in the iteration loop. Without it, weeks slip.

These five habits don’t sound like creative principles. They sound like operations. That’s because they are. AI-driven creative at scale is an operations problem disguised as a creative problem. The teams that solve it look at their creative function as a learning machine, not a production machine.

Campaign structures that make sense for AI-driven creatives

1. Lock the strategy layer before turning on the AI. Hook frameworks, format specs, angle assignments, brand rules. If your AI is generating the strategy, you have no system.
2. Specify visual hooks as themes, not single images. Themes compound across iterations.
3. Use a fixed-format system to make tests comparable. Same handful of formats per concept, every time.
4. Use BANP for scripts, not the training-data favorites. Explicitly forbid PAS, BAB, AIDA in your prompts.
5. Iterate on the eight axes, not just the headline. AI is the production model that finally lets you cover all of them.
6. Run three cohorts in parallel per concept. Hook copy cohort, visual hook cohort, format cohort. Hold everything else fixed.
7. Curate brutally. Selection discipline beats generation volume, every time.
8. Build in operational discipline from day one. Format portability, micro-audiences, naming, themed CPPs, weekly cadence. Without these, the volume becomes noise.

FAQ

What’s the difference between AI-driven creatives and AI-generated creatives?

AI-generated creatives are produced end-to-end by AI tools. The brief is written by AI, the script is written by AI, the visual is generated by AI, and the variant ships without much human review. AI-driven creatives use AI as the production layer only. The brief, the angle, the hook layers, the format, and the curation are still human work. Same production speed, much higher hit rate. Shorthand: AI-generated is what fails. AI-driven is what works.

Does AI creative replace UGC creators entirely?

For performance creative on Meta and TikTok, increasingly yes. For brand work where a real human face carries authenticity that AI faces still lose to, no. As of 2026, faceless visual hooks (mockups, objects, app screens) outperform AI-generated faces on most direct-response placements.

How fast does an AI creative iteration loop actually run?

Concept brief to a published variant runs in under a workday for static formats and under two days for video. UGC ran seven to fourteen days. The bottleneck is no longer production. It’s the strategy lock and the curation step.

What does the team’s time shift to?

Brief writing (more rigorous, more negative-space rules), curation (looking at what generated, killing 90% of it), and analysis (reading test results across more variants than UGC ever supported). Production time goes near zero. Strategic time goes up.

How many variants per concept should I be running?

For a structured concept, the variant tree easily clears 200 permutations across hook copy, visual hook, format, music, VO, and scene order. You don’t ship all 200. Fifteen to twenty in active rotation per concept, across two or three formats, is a reasonable ceiling. The constraint isn’t generation cost. It’s how many variants your ad account can carry without diluting learning.

Can AI generate the strategy too?

No. Or rather, it can, badly. Strategy is the negative space. What the AI is not allowed to do, what hooks are not allowed to look like, what angles the brand can’t credibly own. AI fills the positive space inside the rules. The rules have to come from a human strategist who has shipped enough ads to know where the dead ends are.

Which AI tool should I use to generate visuals?

Depends on the format. Midjourney for stylized stills, Gemini 3 Pro Image for fast realistic scenes, Sora for image-to-video and short clips, Runway Gen-4 for motion control on stills. Don’t pick based on team familiarity. Pick based on what the brief is asking the visual to do.

Do I need a designer if I’m using AI tools?

Yes, more than ever. Designer time shifts from production to composition and editing. Compositing AI-generated elements into a fixed format, color-correcting for brand consistency, fixing the 10% of an output that’s wrong rather than regenerating, polishing typography on overlays. The designer becomes a curator and finisher.

How do I prevent AI from inventing specifics?

In the brief: list every fact the variant is allowed to use, and instruct the AI to use only those. In curation: verify every number, claim, and stat against the source. Treat AI output as draft, not source-of-truth.

What’s the cost difference vs. UGC production?

Production cost per variant collapses by roughly an order of magnitude. The savings get reallocated to strategy, curation, and analysis time. Total program cost moves less than people expect, because the human work shifts rather than disappears. The ROI shows up in iteration speed and learning rate, not in the variable cost line.

Related Reading

• Why fully AI-generated mobile creative keeps losing (and the sequence that fixes it)
• The Sensor Tower finding that makes creative velocity a performance lever
• Why creative diversity matters for ad account health
• How to structure Meta campaigns for creative testing
• The Ultimate Guide to UGC for Performance Marketing

Want the skill pack?

Every prompt in this guide, four Claude Code skills, the BANP and 3C frameworks, and the three-phase curation checklist, in one download.

Get the AI-Driven Creatives Skill Pack on GitHub