Make ads like a pro
A solo operator can now produce production-quality static and video ads with three tools: Claude, Nano Banana 2, and Veo 3.1. What used to be a multi-person, multi-week pipeline fits in an afternoon.
Make ads like a pro
~2,800 words · prompts copy-pasteable · the AI production pipeline used by the best AI-native creative agencies for eight and nine-figure DTC brands
The Tool Stack (And The Cheapest Path)
Will Sartorius's stack: Claude for strategy and prompts, Nano Banana 2 (Google) for static image generation, Veo 3.1 (Google) for video, and optionally fal.ai as a UI on top of Nano Banana for batch generation.
I ran the pricing math. As of May 2026:
| Tool | Daily / monthly access | Cost | Notes |
|---|---|---|---|
| Claude Pro (Anthropic) | Generous daily usage | $20/mo | No native image gen. Great for everything around the image (strategy, copy, prompts, Veo JSON). |
| ChatGPT Plus (OpenAI) | ~50 image prompts per 3-hour window | $20/mo | GPT Image 1.5 included. Fine for ads. Output isn't as clean as Nano Banana 2 for product photography. |
| Gemini app, free tier | 20 Nano Banana 2 images/day | Free | The actual cheapest path. 20/day = ~140/week. For 20-50 ad images/week you have plenty of headroom. |
| Gemini AI Plus | 50 Nano Banana 2 images/day | $19.99/mo | Step up if you're producing high volume or want extra headroom. Same model, more cap. |
| Google AI Studio / API | Pay per image | $0.067-$0.151 per NB2 image | Best when you want programmatic / batch generation. Not the right starter path. |
Cheapest viable stack: Claude Pro ($20/mo) for everything pre-image, plus the Gemini app's free tier for image generation, free until you're producing more than ~140 ads a week (then Gemini AI Plus at $19.99/mo).
Veo 3.1 for video is paid (cents-per-second of output). For animating statics you'll typically generate one 8-second video per concept, so it's a few dollars per ad, which is fine.
The Quick Win Path: Clone Any Competitor's Ad In 5 Steps
This is the lazy-mode workflow. Use it the first week to get your hands dirty and see what NB2 can do.
Deconstruct the competitor ad
Upload a screenshot of the ad to Claude and run the deconstruction prompt. You want a description so thorough that someone who's never seen the ad could rebuild it from text alone.
Prompt: clone deconstruct
I'm uploading a competitor's ad. Deconstruct it completely.
Break down every single element you see:
1. AD FORMAT: What type of ad? (feature callout, headline,
before/after, testimonial, etc.)
2. COPY: Every piece of text — headlines, subheads, product
names, feature labels, callouts, CTAs, fine print. Exact
words and exact position.
3. PRODUCT: What is the product? Angle, orientation, position
in frame, % of frame.
4. LAYOUT: How are elements arranged relative to each other?
What does the eye hit first, second, third?
5. BACKGROUND: Color, gradient, texture, photo — exactly what's
behind the product.
6. TYPOGRAPHY: Font sizes relative to each other, weights,
casing, alignment per element.
7. VISUAL DEVICES: Leader lines, arrows, badges, icons, borders,
shadows — anything decorative.
8. COLOR PALETTE: Dominant colors and how each is used.
9. SPACING: Tight or airy? Centered or asymmetric?
Be thorough. I want someone who has never seen this ad to
understand exactly what it looks like from your description.
Rewrite it for your brand
Same Claude thread. Ask it to convert the deconstruction into a Nano Banana 2 prompt for your product.
Prompt: make it yours
Now rewrite this ad for my brand.
Brand: [BRAND]
Product: [DESCRIPTION, PRICE, GUARANTEE, OFFER, KEY PROOF]
Using the deconstruction above, write me a detailed image
generation prompt for Nano Banana 2 that recreates the same
ad layout and structure but for my product.
The prompt must include:
- Exact dimensions (1080x1920, 9:16 vertical)
- Background treatment (recreate the style but adjusted for
my product's color)
- Product placement: same position, angle, scale as the
original, but describing MY product
- Every piece of copy rewritten for my brand and product —
headline, subhead, every feature callout — same tone,
length, structure as the original
- Typography direction: same size relationships, weights,
casing, alignment
- All visual devices: leader lines, spacing, etc.
- Color palette guidance based on my brand
- A note that I will upload 2-3 of my brand's existing ads
as reference images — tell the generator to match colors,
fonts, visual style from THOSE, not from the competitor
- A note that I will upload a product photo for accurate
product representation
- Safe zones: no critical content in top 270px or bottom 340px
- Rules: all text rendered in the image, no text smaller than
24px, sufficient contrast, one continuous image with no
panels or borders
The prompt should be ready to copy-paste directly into Nano
Banana 2 with no edits needed.
Gather your three uploads
You need: (a) the competitor ad screenshot as layout reference, (b) 2-3 of your existing ads as brand reference, (c) a clean product photo.
Run it in the Gemini app (or AI Studio if you've upgraded)
Open the Gemini app, paste the prompt, upload the three reference images, and generate. The free tier gives you 20 Nano Banana 2 images/day. If you need batch generation or programmatic access, aistudio.google.com exposes the same model via paid API (~$0.067-$0.151/image), and fal.ai is a third-party UI on top.
Pick the winner, ignore the rest
Most won't be usable. That's fine. You're looking for one. Sometimes you'll need to nudge the prompt and re-run. Sometimes the very first generation is the one.
The Quick Win Path: Animate Any Static
The mental model that makes this click: you're not asking AI to animate the ad. You're creating two frames (start + end) and letting AI connect them.
Brainstorm 5 animation concepts
Upload your static to Claude and ask for ideas.
Prompt: animation concepts
I want to animate this static ad. Give me 5 animation concepts.
For each one tell me:
1. Is my static the START or END frame?
2. What does the other frame look like — describe in detail?
3. What motion happens between the two frames?
4. Why does this motion make sense for the product?
Rules: all text must stay frozen in place throughout the
animation, no camera movement, 3-4 seconds max.
Write an NB2 prompt for the missing frame
Pick your favorite concept. Have Claude turn it into an NB2 prompt that generates the other frame.
Prompt: missing frame
I want to animate this static ad. The concept is:
[PASTE CHOSEN ANIMATION CONCEPT FROM STEP 1]
My static is the [START / END] frame.
Write a Nano Banana 2 prompt to generate the other frame as
a still image. Based on the animation concept, figure out
which elements need to be added, removed, or repositioned
in the missing frame.
Every piece of text must be reproduced verbatim — same words,
same position, same fonts, colors, and sizing. The background,
lighting, and overall style must match my static exactly.
Only change what needs to be different for the other frame.
The prompt must specify exact dimensions (1080x1920). I will
upload my original static as a reference image — tell the
generator to match everything from it except the elements
that change between frames.
Be extremely detailed — NB2 won't guess, it only does what
you tell it.
Generate the missing frame in NB2
Same as the static workflow above. Upload the original static as reference + the prompt Claude just wrote.
Write a Veo 3.1 JSON prompt that connects the two frames
Veo is the one tool that requires JSON. Plain text loses you control over timing and motion. Have Claude write the JSON.
Prompt: Veo JSON
I'm uploading two frames for a video animation.
IMAGE 1 is the START frame.
IMAGE 2 is the END frame.
The animation concept: [PASTE CONCEPT]
Write me a comprehensive Veo 3.1 JSON prompt that creates a
smooth 8-second animation from the start frame to the end
frame.
The prompt MUST include:
- Every piece of text from both frames reproduced verbatim
so the fonts don't degrade during generation
- Explicit instruction that ALL text stays completely frozen
in place throughout — no movement, warping, or fading
- No camera movement whatsoever
- The background stays perfectly still
- A detailed description of the motion: what enters, from
where, how it moves, physics and easing
- The animation must resolve cleanly into the end frame as
its final resting state
- Describe what stays still and what moves — be explicit
about both
Format as a ready-to-paste JSON prompt for Veo 3.1.
Run in Veo at labs.google/flow, then iterate
You won't one-shot it. Upload the GIF + both original frames back to Claude and describe what went wrong: "the text warped at 3 seconds," "no gummy bears appeared," "the tube floats up on its own instead of being lifted." Claude rewrites the JSON. Re-run. 2-3 rounds is normal.
The Scalable Path: Build Once, Ship Forever
The quick wins are great, but if you're making 20+ ads a month you want a system that produces consistent on-brand work without you re-writing prompts every time. Will's scalable method has four layers.
| Layer | What it is | Build cadence |
|---|---|---|
| 1. Brand Extraction | A deep-research prompt that scrapes your site, product pages, and competitors to produce a full Brand Bible (.md). | Once per brand. |
| 2. Brand Reference Cards | Two screenshots (Brand Spec Card + Visual Style Card) that you upload to NB2 instead of vague text descriptions of your brand. | Once per brand. |
| 3. Format Templates | Recipe cards (.md) for each ad type: headline, before/after, testimonial, statistics, us-vs-them, etc. | Build 3-5 to start, grow over time. |
| 4. Copy Scoring Agents (optional) | 5-7 .md files that act as automated copy reviewers (persona fit, brand voice, grammar, emotional resonance, format compliance, etc.) | Optional. I think we skip this. |
Build the top row once. Then every new ad is just: pick a format + persona + angle + emotion, Claude writes the brief, you assemble the NB2 prompt with brand cards + product photo attached, generate.
Layer 1: Brand Extraction
This is the longest prompt in the whole system. It's a 7-phase research brief that has Claude search the web, scrape your site, and output structured JSON for each phase: brand identity, product intelligence, proof points, transformation/pain points, competitive positioning, founder story, and a final block of ad-ready hooks for 15 different ad formats.
Will's full version is 900+ words and lives in his prompt library. The short framing:
Prompt: Brand Extraction (condensed)
Target Brand: [BRAND NAME]
Target URL: [URL]
Hero Product: [PRODUCT NAME or "FLAGSHIP"]
Act as a Senior Performance Creative Strategist at a DTC ad
agency. Build a comprehensive Ad Creative Research Brief
covering:
PHASE 1: Brand Identity & Visual System (voice, banned
language, fonts, hex colors, imagery style)
PHASE 2: Product Intelligence (hero claim, how it works,
ingredients, offer, guarantee)
PHASE 3: Proof Points (clinical, press, testimonials, ratings)
PHASE 4: Transformation & Pain Points (before/after states,
emotional triggers)
PHASE 5: Competitive Positioning (top competitors, us-vs-them
angles)
PHASE 6: Founder & Origin Story
PHASE 7: Ad-Ready Hooks for 15 formats (before/after, bullets,
negative marketing, news, handwriting, us-vs-them, statistic,
social proof, press, testimonial, founder, features/benefits,
UGC, carousel, new formats)
Output each phase as structured JSON. Search the web for
"[Brand] brand guidelines pdf", "[Brand] hex colors",
"[Brand] vs [competitor]", "[Brand] founder interview", etc.
Save the final output as brand-bible.md
Layer 2: Brand Reference Cards
Will's argument: instead of uploading 2-3 reference images and hoping NB2 figures out your aesthetic, you generate two purpose-built reference cards as screenshots that explicitly lay out your brand. NB2 reads them and uses them as ground truth.
Prompt: Brand Spec Card
Using the Brand Research Brief you just created, generate a
clean HTML page I can screenshot as a Brand Spec Card for AI
image generation.
Include:
1. Logo & Wordmark — show the brand name in the exact fonts
and colors from the brief, with black and white versions
2. Typography System — sample text rendered in each font role
(headlines, body, accents/labels) with font names labeled
3. Color Palette — visual swatches with hex codes (primary,
secondary, accent, background)
4. Design Rules — 4-5 "always do" and 4-5 "never do" rules
5. CTA Button Style — render the actual button (color, shape,
text treatment)
Make it clean and visually explicit. This will be uploaded
as a reference image to NB2, so every detail must be visual,
not just described in text.
Once complete, render it so I can screenshot at high resolution.
Prompt: Visual Style Card
Generate a second HTML page as a Visual Style Card.
Include:
1. Brand Essence — 3 adjectives that define the brand voice
2. Founder Quote — one defining quote
3. Photography Direction — visual blocks for: product
photography, model photography, lifestyle/context, and
background/surface preferences, each with keyword tags
4. Always/Never Rules — specific to how the brand portrays
people, skin, products, environments
5. Product Styling Notes — how the hero product should be
photographed (open vs closed, texture visible, etc.)
6. Mood Spectrum — where the brand sits on scales like loud
vs quiet, youthful vs timeless, clinical vs warm
Same as before: clean, visual, designed to be uploaded as
a reference image.
Layer 3: Format Templates
For each ad format you ship regularly (headline, before/after, testimonial, statistic, us-vs-them, etc.), build a .md "recipe card" that defines:
- What the format is in one sentence (e.g., "A single dominant headline carries the entire message.")
- What copy goes where (headline / subhead / CTA slots)
- Word counts or character constraints per slot
- Image direction: composition, lighting, product position, background
- What to avoid (generic imagery, text over busy areas, too many elements)
- Safe zones
- What must never change vs. what can vary across versions
Fastest way to build them: collect 5-100 screenshots of ads in the format, upload to Claude, and ask:
Prompt: format template extraction
Analyze these ads. What do they have in common?
Write me a .md file with copy slots, word counts, image
direction, layout rules, and "always / never" guardrails.
Include safe zones, banned words, element limits, and what
to avoid. Output as headline-format.md (or whatever this
format is).
Then manually edit it with your specific brand guardrails. Test it. Iterate. Build 3-5 templates to start (headline, before/after, testimonial, social proof, us-vs-them is a great Starter 5).
Layer 4: Copy Scoring Agents (skip for now)
Will builds 5-7 .md "agents" that score each brief 0-100 on persona fit, brand voice, grammar, and the like, with nothing shipping under 90. At agency scale (50+ ads a week) that gate earns its keep. Below ~30 briefs a week, just read each brief yourself. When you do need it, start with one combined QC agent, not seven.
Putting It Together: Briefing And Generating One Ad
Assume the four layers are built. Here's the start-to-finish for one ad:
Prompt: write the brief
I want to create a [FORMAT] ad for [PRODUCT].
Persona: [PERSONA]
Angle: [ANGLE]
Emotion: [EMOTION]
Use my Brand Bible and Format Template to write the full
brief: headline, subhead, body copy, CTA, creative direction
— everything. Follow the format template exactly.
Once the brief is approved (read it yourself, edit it, ship when you're happy, see the Layer 4 note above):
Prompt: convert brief to NB2 prompt
Convert this into a ready-to-paste Nano Banana 2 image
generation prompt.
The prompt must include:
- Exact dimensions (1080x1920, 9:16 vertical)
- Every piece of approved copy rendered verbatim in the image
- Product position, angle, scale from the creative direction
- Background treatment from the creative direction
- Typography: exact font names, weights, colors from the
Brand Spec Card
- Safe zones (top 270px, bottom 340px clear of critical
content)
- Lighting and energy direction
- Logo placement per Brand Spec Card
- Universal rules: no panels, no text under 24px, sufficient
contrast, photorealistic product
- A note that I will upload the Brand Spec Card, Visual Style
Card, and product photo as reference images
Then upload everything (NB2 prompt + brand cards + product photo) to the Gemini app (free tier) or AI Studio (if you've upgraded to paid). Generate 4-8. Pick the best.
Multiplying Winners Across Formats
Once one concept lands, don't re-brief from scratch for every other format. Use the same persona/angle/emotion/proof and let Claude repackage it.
Prompt: multiply formats
I have a winning ad brief that I want to multiply across
other format templates. The original brief is below.
Here are my format templates: [list or upload your .md template files]
For EACH format template, rewrite the brief to fit that
format exactly:
- Keep the same persona, angle, emotion, and core product
truth
- Rewrite the copy to match the new format's structure
(headline lengths, copy slots, required elements)
- Follow the format template's copy rules and variation vectors
- Adjust the creative direction to match the new format's
visual requirements
- Write a complete Nano Banana 2 image generation prompt
for each
Do not change the strategic foundation. Only change the
packaging.
Original brief:
[PASTE WINNING BRIEF]
You now have one strategically validated concept turned into 5 ads in 5 formats, ready to test against each other in Meta.
Common Failure Points
| If this happens | Likely cause | Fix |
|---|---|---|
| Image looks off-brand | Weak Brand Extraction or Reference Cards | Add exact hex colors, font names, logo rules, image style, do/don'ts. |
| Layout shifts feel random | Format template too loose | Define hierarchy, spacing, safe zones, copy placement, what must not change. |
| Copy feels generic | Persona / angle too vague | Add persona language, trigger moment, specific proof. |
| Product is warped, text changes | NB2 prompt too short | Lock product, text, proportions, camera, colors explicitly. Aim for 1,000+ words. |
| Animation is messy | Frame pair or Veo JSON underspecified | Rebuild the frame pair, simplify the motion, tighten JSON constraints. |
Sources and credits. The AI creative pipeline, clone-and-rewrite workflow, and frame-pair animation method in this guide are taught publicly by Will Sartorius, CEO of SelfMade / Skipper, including this video. We're crediting him for teaching and refining these frameworks, not necessarily for inventing every component. The sequencing into one operating system, the simplified pipeline for solo operators, the current pricing math (verified May 2026), and the starter-vs-scale guidance are ours. If the frameworks land, that's because of Will. If anything is wrong, that's on us.