Ad Image Prompt Library

ad-image-prompt-library

A library of copy-paste ad-creative recipes, in two families:

• Part A — HTML-mockup ad formats. Cheap, pixel-exact, deterministic. Render as HTML/CSS → headless screenshot, no image/video model. Every character correct, every hex exact. Build these with the ad-creative-templates pipeline (one template + a data row → many variants).
• Part B — AI-image-gen prompt templates. Reusable prompt skeletons for a gpt-image / Nano-Banana / GPT-Image-style model. Run these through the ad-creative-generation mechanics (endpoint choice, reference-image rules, vision-QA loop).

Everything here is generic — swap the [PLACEHOLDERS] for your product, palette, and copy. Truthfulness still applies: no fabricated quotes/reviews/ratings, no impersonation of a real person or platform, only real features (see ad-creative-templates §6).

Which format for which goal (selector)

Rule of thumb: if the creative is text-, UI-, or logo-heavy, it's Part A (a model garbles it). If it's a photoreal scene or a stylized character, it's Part B.

PART A — HTML-mockup ad formats ("iOS Ads" family)

These are the highest-performing cheap format: they render pixel-exact from HTML/CSS, cost ~nothing per variant, and win because —

• Low cognitive load — people read these exact UIs (Messages, Notes, ChatGPT, WhatsApp, the lock screen) dozens of times a day; the brain parses them with zero effort.
• Native / non-ad feel — they look like a screenshot a friend sent, not a branded ad, so they slip past ad-blindness.
• Curiosity — watching someone type (typing indicator, a half-written note, a streaming answer) pulls the eye and holds it to the reveal.

Reproduce the platform chrome faithfully — a wrong font, a fake battery icon, or off spacing breaks the "this is real" illusion instantly. Render each at feed 1080×1080 and story 1080×1920 (see ad-creative-templates §4). Below, each format lists: what it is · why it performs · key UI to reproduce · fill-in template.

A1 — iMessage chat-reveal

• What it is: a fake iMessage thread where a short conversation reveals the product — a question/pain, a recommendation, and a tappable product link card.
• Why it performs: the most-read UI on iOS; blue/grey bubbles read as a real text from a friend recommending something → maximum native feel + social proof by proxy.
• Key UI to reproduce faithfully:
  • Contact header: back chevron, centered contact avatar + [NAME], top status bar (time, signal, wifi, battery).
  • Bubbles: incoming = grey left-aligned, outgoing = iMessage blue (#3B7DFF-ish gradient) right-aligned, rounded ~18px with the little tail on the last of a run.
  • Optional typing indicator (three grey animated dots in a bubble) as the last incoming — the curiosity beat.
  • A link-preview card (rounded, thumbnail + title + domain) for the product URL.
  • Bottom iMessage input bar: rounded "iMessage" field, camera/plus icon left, a grey ↑ send arrow right.
• Fill-in template:

  Header: [FRIEND_NAME]
  Grey (them): "[PAIN / QUESTION, e.g. ugh still can't find a [CATEGORY] that [DESIRED OUTCOME]]"
  Blue (you):  "omg just use [PRODUCT]"
  Blue (you):  "[ONE-LINE WHY — real benefit]"
  Link card:   [THUMBNAIL] · "[PRODUCT] — [TAGLINE]" · [DOMAIN]
  Grey (them): "wait this is exactly what I needed 😍"
  [typing indicator on them, optional]
  

A2 — Apple Notes reveal

• What it is: a single Apple Notes note titled like a listicle hook ("Things nobody told me about [X]"), with a few bullet lines, a blinking cursor, and the Notes formatting toolbar + keyboard visible.
• Why it performs: reads as a private, honest jot — not marketing. The listicle title promises value; the visible cursor/keyboard says "being written right now," which is the curiosity hook.
• Key UI to reproduce faithfully:
  • Notes header: back to folders (yellow chevron Notes), share icon, compose icon; date/time line under the title in grey.
  • Title line bold + larger; body in the Notes body font; a text cursor (thin blue I-beam) at the current line.
  • Formatting toolbar row above the keyboard (Aa format, checklist, camera, pencil/markup).
  • The iOS keyboard at the bottom (QWERTY, spacebar, blue return) — or the predictive-text strip alone for a tighter crop.
• Fill-in template:

  Title: Things nobody told me about [TOPIC / PAIN]
  [DATE • TIME]
  • [MYTH or mistake people make]
  • [The thing that actually worked → mentions [PRODUCT] naturally]
  • [Concrete result / number, real]
  • [cursor here|]
  

A3 — ChatGPT chat-reveal

• What it is: a mock ChatGPT thread — a user question in a grey bubble, then a "streaming" assistant answer with a short intro line and bullets that recommend/explain the product.
• Why it performs: AI answers are read as neutral, authoritative "research," so a recommendation inside one carries borrowed credibility; the mid-stream cursor sells "it's answering right now."
• Key UI to reproduce faithfully:
  • Top bar: ChatGPT title (or model name), new-chat/menu icons.
  • User turn: right/grey rounded bubble with the question (ChatGPT web shows user messages in a light bubble; mobile similar).
  • Assistant turn: the ✦/logo avatar, left-aligned answer, a one-line intro then a markdown bullet list; optional bold lead-ins per bullet.
  • A streaming caret (▍) at the end of the last line, or a "Stop generating" pill, for the live feel.
  • Bottom composer: rounded "Ask anything" field, mic + send icons.
• Fill-in template:

  User (grey): "best [CATEGORY] for [SPECIFIC USE-CASE / CONSTRAINT]?"
  Assistant:
    Great question — here are the top options:
    • **[PRODUCT]** — [why it fits the constraint, real].
    • [Honest alt / criterion] ...
    • [Honest alt / criterion] ...
    For [USE-CASE], [PRODUCT] is usually the best starting point.▍
  

A4 — WhatsApp chat-style

• What it is: a WhatsApp thread themed around a "breakup with the pain" bit — the pain is personified and "leaves the chat," with a product image shared in the thread.
• Why it performs: the green WhatsApp chrome is instantly familiar; the "[Pain] has left the chat" system line is a native meme format that reads as funny/organic, not salesy.
• Key UI to reproduce faithfully:
  • Header: green (#075E54 / #128C7E) bar, back arrow, contact avatar + [NAME], and the "online" / "last seen today at [TIME]" subtitle.
  • Chat background = WhatsApp's faint doodle wallpaper; bubbles: incoming white, outgoing light-green (#DCF8C6), with a tiny time + double blue tick on outgoing.
  • A shared product image bubble (thumbnail + caption).
  • A centered grey system message pill: "[PAIN] has left the chat" / "[PAIN] was removed".
  • Bottom type-a-message bar: rounded field, emoji + attach + camera icons, a green mic/send button.
• Fill-in template:

  Header: [PRODUCT or FRIEND] — online
  White (them): "still dealing with [PAIN]?"
  Green (you):  "not anymore lol"
  Green (you):  [product image] "[PRODUCT] fixed it — [real benefit]"
  — system —   "[PAIN] has left the chat"
  White (them): "sending this to everyone 😭"
  

A5 — iPhone lock-screen notification

• What it is: an iPhone lock screen — big clock + date over a scenic wallpaper — with a stack of notification banners (each with an app icon) that tell a micro pain→relief story. Shot to look like a raw screenshot.
• Why it performs: the single most-seen screen on earth; a notification is glanceable in <1s and feels like a real alert, not an ad. Stacked banners let you sequence a tiny narrative.
• Key UI to reproduce faithfully:
  • Time (huge, SF font) + date line above it; lock icon at top; flashlight + camera buttons bottom corners (optional).
  • A scenic wallpaper (mountain/beach/gradient) — keep it real-photo, slightly soft.
  • Notification bubbles: frosted-glass rounded rectangles, each = rounded app icon + bold [APP NAME] + relative time (now / 2m ago) + one/two lines of message text; stack 2–3, newest on top, slight overlap on the bottom one.
  • Raw-screenshot look: real status bar (carrier, 5G, battery %), no drop-shadow "ad polish," no added frame.
• Fill-in template:

  9:41   Monday, [DATE]
  [wallpaper: scenic photo]
  [🔔 app icon] [PRODUCT]        now
     "[Push copy: benefit / social proof, e.g. '[N] people started [OUTCOME] today']"
  [💬 app icon] [FRIEND]         2m ago
     "have you tried [PRODUCT]?? game changer"
  [📅 app icon] Reminders        8m ago
     "Stop [PAIN] — try [PRODUCT]"
  

PART B — AI-image-gen prompt templates

Skeletons for a text→image / image-edit model. Fill the [placeholders], then run through ad-creative-generation (pick generations vs edits, index reference images, vision-QA every render). Do not ask these for exact headlines, logos, or real UI — that's Part A.

Universal realism levers

• SLCT — every prompt fills four slots, no free-associating: Subject (who/what, exact count), Lighting (source, direction, quality — "soft window light from left"), Camera (lens + framing — "35mm, eye-level, medium close-up"), Technical (medium, film/sensor look, texture — "shot on phone, visible grain, real skin pores"). Naming the lens + light is what tips a render from "AI" to "photo."
• Text-in-quotes = literal on-image text. Put any words you actually want rendered in double quotes in the prompt (the sign reads "[EXACT TEXT]"). Keep it to a few big words — models still garble long strings, so anything precise belongs in a Part-A template.
• Identity-lock (when using a reference photo of a person/product): state keep the exact same face/identity/product as the reference — do not restyle, age, slim, or beautify; same features, same proportions. Prevents the model drifting the subject into a generic look-alike.
• Negative-prompt hygiene: use negatives for style/medium/defects, not for content ("no content" is unreliable). A reusable baseline: no extra fingers, no fused hands, no warped face, no garbled text, no watermark, no logo, no studio polish, no plastic skin, no oversaturation, no duplicated limbs.

B1 — Photo-of-a-screen hyper-realism

• When to use: you want a "someone snapped a photo of their phone/laptop showing [SCENE]" look — the anti-polish, maximally-believable UGC vibe (also the base for faking a screenshot inside a real environment).
• Prompt skeleton (JSON-style keeps the model on-spec):

  {
    "shot": "this is a PHOTO OF A SCREEN, not a screenshot — a phone/monitor physically photographed by another camera",
    "subject": "[DEVICE] displaying [SCENE / CONTENT]",
    "screen_realism": ["visible RGB sub-pixel grid", "faint moiré banding", "screen glare / reflection", "dust specks and fingerprint smudges on the glass", "slight off-axis angle", "backlight glow at edges"],
    "environment": "[real setting — desk, hand-held, café table], ambient [LIGHT]",
    "camera": "[35mm / phone camera], [framing], natural handheld imperfection",
    "technical": "real photo, natural grain, NO studio polish, NO perfect crop",
    "identity_lock": "[if a product/person is shown: keep it exactly as reference]",
    "negative": "no crisp digital screenshot, no perfect pixels, no clean UI export, no vector sharpness, no studio lighting, no garbled text"
  }
  

B2 — Instagram pop-out / 3D frame-break

• When to use: scroll-stopping "subject bursts out of the post" gimmick for a feed placement.
• Prompt skeleton:

  A centered white Instagram post frame on a plain [BG COLOR] background, with the standard IG UI: heart / comment / share row, "[N] likes", username "[HANDLE]", and caption "[SHORT CAPTION]".
  [SUBJECT] is photographed inside the post but POPS OUT — [head / hand / product] breaks past the white border into 3D, casting a soft drop shadow onto the frame; a hand grips the edge of the frame.
  Lighting: [soft studio / directional]. Camera: [50mm, slight low angle]. Technical: photoreal, crisp subject, realistic shadow + depth.
  Negative: no warped hands on the frame edge, no garbled UI text, no duplicated icons.
  

B3 — Optical-illusion "stepping out of the phone"

• When to use: surreal attention-grab — the subject is literally climbing out of a phone screen.
• Prompt skeleton:

  A large [PHONE] lying/standing on [SURFACE]. [SUBJECT]'s torso emerges UP and OUT of the screen in full 3D, while their legs are still flat ON the screen showing the phone's camera-app UI (shutter button, mode strip) around them — a seamless optical illusion.
  Handwritten annotations / arrows around the scene reading "[SHORT LABEL]".
  Lighting [L], camera [C, slight top-down], photoreal composite, believable shadow where body meets glass.
  Negative: no melted body, no broken perspective, no garbled UI, no extra limbs.
  

B4 — Cinematic character board

• When to use: a premium "cast sheet" for a persona, mascot, or feature line-up.
• Prompt skeleton:

  A dark premium character sheet with thin gold borders, laid out as panels.
  Hero panel: [CHARACTER] full-body, [pose], dramatic [L] lighting.
  Surrounding smaller panels: multiple views (front / 3-4 / profile / back) and 3-4 facial EXPRESSIONS (neutral, smiling, focused, surprised).
  A vertical PALETTE STRIP of the character's key colors down one side.
  Small handwritten-style labels "[NAME]", "[TRAIT]". Style: cinematic, cohesive, high-detail concept-art.
  Negative: inconsistent character across panels, garbled labels, mismatched colors.
  

B5 — Isometric figurine

• When to use: a cute collectible-toy render of a person/mascot — high shareability.
• Prompt skeleton:

  A miniature full-body ISOMETRIC realistic figurine of [SUBJECT / this person], [outfit / props], standing on a small round base, on a plain white background, soft even studio light, sharp focus, 4K, product-render look.
  [identity-lock if from a reference photo]
  Negative: no full-size human, no busy background, no warped face, no extra base clutter.
  

B6 — OOTD / mirror-selfie Pinterest collage

• When to use: aesthetic lifestyle / fashion / mood board that reads as organic Pinterest content.
• Prompt skeleton:

  A 4-image collage (2x2), Pinterest / moodboard aesthetic: [SUBJECT] in [OUTFIT / SETTING] — one mirror-selfie holding a phone, one detail shot, one full-body, one candid.
  Overlay a music-player widget (album art "[TRACK]", play bar, artist "[ARTIST]") on one tile.
  Look: desaturated muted tones, visible film grain, soft natural light, [palette]. Cohesive edit across all four.
  Negative: oversaturation, mismatched lighting between tiles, garbled widget text, warped hands on the phone.
  

B7 — Anime / GTA-style poster grid

• When to use: a stylized fandom/hype poster — a hero center panel ringed by vignette panels.
• Prompt skeleton:

  A [anime cel-shaded / GTA loading-screen illustrated] poster grid: one large CENTER panel of [HERO SUBJECT / scene], surrounded by smaller vignette panels (side characters, objects, locations, close-ups).
  Bold black gutters between panels; a [PALETTE] color scheme with a palette strip along the bottom edge.
  Optional title text "[TITLE]" in the [style]'s signature lettering.
  Style: high-contrast, dynamic, cohesive [anime / GTA] art direction.
  Negative: muddy panels, garbled title, inconsistent art style between panels.
  

Anti-patterns

• Asking a Part-B model for the Part-A stuff — an iMessage thread, a lock-screen, exact push copy, a real logo. It garbles the chrome and the text. Those are HTML mockups.
• Sloppy platform chrome in Part A — a fake battery icon, wrong bubble color, missing typing indicator, off font. The whole edge is "this looks real"; a tell kills it.
• Long literal text in a Part-B prompt. Quotes help for a few big words; anything precise still drifts — template it.
• Skipping identity-lock when a real person/product is the reference — the model drifts them into a generic look-alike.
• Fabricated social proof. A fake testimonial in an iMessage/ChatGPT reveal, invented likes/reviews, impersonating a real person or platform. Real data only (ad-creative-templates §6).
• Shipping without QA. Part A: eyeball the chrome at 1:1. Part B: run the vision-QA loop (ad-creative-generation §5).

Pairs with: ad-creative-templates (the HTML→screenshot pipeline that renders every Part-A format, + the truthfulness rule), ad-creative-generation (image-model endpoints, reference-image rules, and the vision-QA loop for every Part-B render), ad-experiments (which format/angle to test and how to judge the result), high-fidelity-ui-image-gen (the gen-vs-screenshot fidelity tradeoff), app-screenshots (capturing real product UI to drop into a mockup).