Your first ad — the deep dive

Reading time: 15 minutes. Hands-on time: 20 minutes. What you'll have at the end: A generated ad you understand line by line. The mental model for picking the right concept type for any product. The vocabulary to direct the AI when an output isn't right.

This tutorial assumes you've finished 01, 02 — Shopify, and ideally 03 — Meta (though you can generate ads without Meta connected — you'll just download instead of publish).

What an ad is in Scarif One

An ad is one piece of copy + one image generated together so they reinforce each other. The copy informs the image prompt; the image informs the copy's emotional register. Generating them separately produces incoherent ads — that's why we always generate both at once.

Each ad has:

Headline — 6-12 words, the hook
Primary text — 30-150 chars, expands the hook
Image — square 1080×1080, vertical 1080×1350, or story 1080×1920 depending on placement
Concept type — the "shape" of the ad. There are 9 shapes, each with a distinct visual + copy pattern (more below).

The 9 concept types

Every ad fits into one of these. Pick the right one and the AI does 80% of the work for you. Pick the wrong one and even great copy will look mismatched against the image.

1. Hero

The product, perfectly lit, against a clean backdrop. Classic e-commerce ad. Use when: you have a beautiful product photo and want it front and centre. Best for: premium goods, food, fashion.

Example copy register: "Made for the morning person who refuses to compromise." → big, simple visual.

2. Founder-direct

The founder talks to camera. AI generates a portrait-style image (or uses one you upload from your asset library) plus copy in first person. Use when: your brand has a face and a story. The single highest-converting concept type for new brands.

Example: "Hi, I'm Sarah. Three years ago I quit my job to make this rum. Try a bottle — if you don't love it, I'll refund you personally."

3. Ingredient close-up

Macro-style shot of one ingredient (e.g. coffee beans, cocoa nibs, gin botanicals). Copy talks craft and sourcing. Use when: your product story is about quality of inputs.

4. In-context

The product being used in real life — coffee in hand on a morning walk, gin being poured at a dinner party. Use when: lifestyle matters more than the product itself. Most flexible concept type.

5. UGC quote

A real customer quote (pulled from your Judge.me reviews) overlaid on a customer-supplied photo (also from Judge.me UGC). Use when: you have 5+ five-star photo reviews. Highest-converting for ads pointed at lookalike audiences.

6. Founder direct (text-only)

A handwritten-feel note from the founder against a colour-block background. Use when: you don't have a great founder photo but the message is personal.

7. Landmark backdrop

The product in front of a local landmark (e.g. coffee in front of the Brighton Pavilion). Requires you to fill in localLandmarks during setup. Use when: you serve a local market and want the cultural specificity.

8. Process shot

Behind-the-scenes — the machine that roasts the coffee, the still that distils the rum, the kiln that fires the ceramics. Use when: craft is the differentiator and the process is visually interesting.

9. Customer quote (text-driven)

A pull-quote from a review, big and bold, with minimal visual. Like a print-ad headline. Use when: the review itself is the entire pitch.

Step 1 — Open the generator

In the dashboard, click 🎯 Ad Creatives (top-left tile) or open /create/ad-creatives directly.

You'll see a 2-column layout:

Left: the configuration form
Right: the gallery of generated ads (empty if this is your first time)

Step 2 — Configure your first ad

Fill in each field. Take 30 seconds per field — better inputs save you regenerations later.

Field: Product

Dropdown of products synced from Shopify (or "Create generic" if you skipped Shopify).

For your first ad, pick your hero product — your bestseller. The one you'd advertise if you had only one. AI generates better copy when the product has a strong identity.

Each product shows a thumbnail and the AI-generated creative brief. Hover over any product to read the brief — that's exactly what the generator gets as context.

No products listed? Either Shopify isn't connected (re-do Tutorial 02), or the first-sync is still running (give it 2 minutes), or you have no active products in Shopify.

Field: Concept type

Pick from the 9 types above. For your first ad, pick Hero or Founder-direct — they're the most forgiving and visually consistent.

If you want to play, generate the same product as Hero and Founder-direct back-to-back. Compare. You'll learn more from that comparison than from any tutorial.

Field: Audience

One short line. The AI uses this to set the copy's register and the image's casting/setting.

Good audience descriptions:

"Specialty-coffee drinkers who roast at home and won't drink supermarket coffee"
"First-time gin buyers, mid-30s, looking for something to give as a gift"
"Loyal customers who've bought 3+ times — we want to thank them"

Bad (too generic):

"Coffee drinkers"
"Anyone who likes gin"
"Customers"

Specificity is the entire game. Spend 60 seconds writing a good audience line.

Field: Aspect ratio

Pick the ad placement:

1:1 (1080×1080) — feed posts on FB + IG. Default.
4:5 (1080×1350) — vertical feed, takes more screen real estate, ~10% better engagement
9:16 (1080×1920) — Story / Reel placement

For your first ad, pick 1:1 — most flexible. You can re-render the same concept in 4:5 or 9:16 later from the gallery card.

Field: AI quality (the toggle)

Bottom of the form. Two options:

⚡ Fast (Flash) — Gemini 2.5 Flash. Generates in 8-15 seconds. Cost: ~£0.04. Default.
🎯 Best (Pro) — Gemini 2.5 Pro. Generates in 30-50 seconds. Cost: ~£0.20. ~5× the cost.

Default: Fast. Use Best only when:

The Fast output isn't quite right after 2-3 regenerations
You're generating a hero ad for a high-stakes campaign
The product has a complex story Pro will narrate better

The toggle remembers your choice across sessions. Set it to Fast and forget it.

Step 3 — Click Generate

Two things happen in this exact order:

3a. The approval gate fires

A pulsing ⏸ button appears in the bottom-right corner of the screen. The system has paused before spending tokens. No AI cost has been incurred yet.

Click the widget. You see:

🤖 Pending approval

Generation: Hero ad — Acme Espresso Beans (1:1)
Estimated cost:  ~£0.04 (Flash copy + Flash Image)
                  £0.21 against this month's £25 cap
What happens:    1 ad copy + 1 image at 1080×1080

[Approve]  [Reject + give feedback]  [Approve & don't ask again for this feature]

This gate fires for every AI-spend action until you toggle it off per-feature in /approvals. It's annoying for the first 5 generations and lifesaving every time something goes wrong (a runaway script, a UI bug, a misclick).

Recommendation for your first 5 ads: keep the gate on. After that, open /approvals and toggle "auto-approve text-only generations" but keep "image generations" gated. 80% of cost is in images; gating those is the right place.

Click Approve. The widget collapses to a green ✓ for 2 seconds.

3b. Generation starts

Right side of the screen, a new card appears with a progress indicator:

Copy generating… (~3 seconds for Flash, ~12 for Pro)
Image generating… (~10-15 seconds for Flash, ~25-40 for Pro)
Done.

The card now shows your headline, primary text, and image.

Step 4 — Read the output critically

Don't accept the first generation reflexively. Read it like a sceptical editor:

Check the copy

Headline test: Does it grab in 3 seconds? Out loud, read it like a friend showed it to you. Does it sound like you, or generic AI-speak?

Primary text test: Does it earn its space? Could you cut a sentence and lose nothing? If yes, regenerate.

Specificity test: Does it mention something only your brand would say? Or is it interchangeable with any competitor's ad?

Check the image

Brand-fit test: Does the colour palette match your brand? Lighting feel? Composition?

Product accuracy test (CRUCIAL): Look at the product in the image. Is it actually your product? AI image models hallucinate — it might've added an ingredient that isn't in your real product, or changed the bottle shape.

Trust test: Would your customers look at this and think "yeah, that's the brand I trust"?

Step 5 — When the output isn't right

You have four moves, in order of effort:

Move 1: Regenerate (free for retry; ~£0.04 again)

Click ↻ Regenerate on the card. Same configuration, fresh AI run. Often the second attempt is better. Three regenerations max — if it's still wrong, the issue isn't the AI, it's the input.

Move 2: Regenerate with feedback

Click ↻ Regenerate with feedback. A textarea opens. Tell the AI what was wrong:

"The image had foam in the espresso cup — Acme Espresso is served straight, never with foam. Also the headline is too generic — make it specific to morning coffee rituals."

The feedback gets saved as a persistent constraint on the product. Every future generation of Acme Espresso obeys this rule. You're teaching the system, one regeneration at a time.

Move 3: Edit the brand voice

If multiple ads sound generic, the issue is upstream — your brand voice doc isn't specific enough. Open /setup → Brand voice → expand the document. Add:

5 specific phrases you'd use ("morning drinkers", "the slow pour", etc.)
3 phrases you'd never use ("artisanal", "synergy", "premium experience")
One paragraph contrasting yourself against a named competitor

Save. Regenerate. The output should be noticeably more "you."

Move 4: Edit the product creative brief

If outputs misrepresent the product itself, the issue is the AI brief. Open /products → click the product → scroll to "Creative brief" → click Edit.

Edit the brief directly. Especially:

Visual identity ("amber liquid in a clear bottle, dark label") — this drives the image generator
Hard constraints ("never show with foam", "never include lemon", "never call it 'craft'")

Save. Regenerate. Ad should now respect your edits.

Step 6 — Save (or publish, or download)

Once you have an ad you'd ship:

If Meta is connected

Click Publish to Meta → opens a publish modal → schedule, post immediately, or save as draft to Ads Manager.
Performance data flows back into Scarif One within an hour.

If Meta isn't connected

Click Download zip — gets you the full creative bundle (image PNG, copy in .txt, Meta-compliant .csv for upload).
Upload manually in Meta Ads Manager / TikTok Ads / wherever.

Either way

The ad is saved in your tenant's gallery permanently. You can re-render it for different aspect ratios, regenerate variants, or export later.

What good looks like (mental model)

After 5-10 generations you'll start to feel which ads are "good." Until then, three rough heuristics:

Would you screenshot it for inspiration? If you saw this ad in your competitor's feed and screenshotted it, that's a good ad. If you'd scroll past, regenerate.
Would your most picky friend share it? Imagine your friend who hates marketing. Show them the ad. If they'd say "actually this is okay," that's a good ad. If they'd say "ugh, more AI slop," regenerate.
Does it still work without the image? Read just the copy. If it stands alone as a tweet or a one-line text, it's a real ad. If it falls apart without the image, the copy is decoration.

Cost rule of thumb

Your first 50 ads will cost you ~£2 in AI fees if you stay on Fast.

Past that, a typical small business generates 50-200 ads/month and spends £5-25/mo on Gemini API. Open /admin/health weekly to see the actual number for your account.

The spend cap (/spend-caps) is a hard ceiling. It's mathematically impossible to exceed — every AI call passes through a budget check before hitting the API. If you set it to £15/mo and try to generate ad #51 in a busy month, the system pauses and asks you to either raise the cap or wait until the calendar rolls.

Where to next

Tutorial 09 — Spend caps + the approval gate (now that you've seen them in action)
Tutorial 06 — Drafting email campaigns (same pattern, different output)
Tutorial 14 — Personas — pre-flight test ads against synthetic customers before publishing
Tutorial 13 — Setup Concierge — the autonomous agent that proposes a week of content for you