Prompt writing is the highest-leverage skill in generative ad creative. A weak prompt produces a mediocre ad that needs hours of cleanup. A strong prompt produces a campaign-ready asset in the first generation. After cataloging the prompts behind several thousand of our top-performing ads, four patterns show up over and over. This post is the working tour of those patterns with concrete before-and-after examples.

The teaching style most online prompt guides use — "be specific!" "use adjectives!" — is true but unhelpful. The shift that actually changes your output quality is moving from natural-language briefs to structured prompts with named slots. Once you internalize a few slot patterns, prompt-writing stops being a guessing game and becomes something closer to fill-in-the-blank composition.

Pattern 1: The five-slot specification

This is the workhorse pattern and the one we recommend every team start with. Every prompt fills five slots in this order: subject, setting, lighting, composition, style. Skip any slot and the model fills it with a default that may or may not match your intent.

Weak prompt: "A nice photo of our skincare bottle"

Strong prompt: "Subject: a 30ml amber glass serum bottle with a black pump cap. Setting: on a smooth white marble surface, fresh eucalyptus leaves to the right. Lighting: soft directional light from the upper left, gentle highlight on the curve of the bottle. Composition: three-quarter angle, bottle filling the center third of frame, shallow depth of field. Style: editorial product photography, premium beauty magazine aesthetic."

The strong version is roughly five times longer and produces dramatically more usable outputs. In our internal tests, the strong format generates a first-try usable asset in eight out of ten attempts. The weak format hits maybe three out of ten and the variance between attempts is much higher.

The slot ordering matters. Subject first because the model anchors most strongly on the first noun. Style last because style cues are easier for the model to weight if they come after the concrete elements. Reversing the order — leading with "editorial style photography of..." — produces noticeably less consistent results.

Pattern 2: Negative space prompting

This is the pattern that fixes the "model put something weird in the background" problem. Image models tend to fill frames densely because their training data was full of dense, busy stock photography. If you want a clean composition you have to explicitly ask for it.

Weak prompt: "Minimalist product shot of running shoe"

Strong prompt: "Product shot of a single running shoe centered in frame. No text. No background patterns. No additional products. No people. Solid gradient background only. Negative space dominates the upper two-thirds of the frame."

The trick is to enumerate the things you don't want. "Minimalist" alone is ambiguous to the model — it might interpret that as "Scandinavian minimalist" with wood textures and house plants, which is not what you want. "No additional products. No backgrounds patterns. No people." is unambiguous. The model defers to explicit exclusions.

The strong version produces clean compositions roughly nine out of ten times. The weak version produces something with at least one extraneous element about half the time.

Pattern 3: Reference-image anchoring

This is the pattern that unlocks brand-consistent work at volume. Instead of describing your product in words, upload a reference photo and tell the model to use it as the source of truth for the product itself.

Weak prompt: "Our product in a kitchen setting" (no reference attached)

Strong prompt: "Setting: a sunlit kitchen with white subway tile backsplash and a wooden cutting board in the foreground. The product should match the attached reference photo exactly in shape, label position, and color. Lighting: warm morning light from a window on the right." (with a clean, well-lit reference photo attached)

The output quality difference is the difference between an ad you can ship and an ad you cannot. Without the reference, the model generates a "kitchen product" — vaguely yours, vaguely not. With the reference, the model generates a kitchen scene with your specific product, label readable, proportions correct, across dozens of variants.

One important rule: the reference image needs to be clean. A reference photo with a messy background or unflattering lighting transfers those qualities to the output. Spend the time to shoot one perfect reference photo per SKU and use it for everything. The investment pays back roughly forever.

Pattern 4: The cinematic camera spec

This pattern is specific to video generation but transforms the output quality when you use it. Instead of asking for "a video of someone using the product," specify the camera as if you were briefing a director of photography.

Weak video prompt: "8 seconds of a person using our face serum"

Strong video prompt: "8 seconds. Shot: medium close-up on a woman in her early thirties at a bathroom vanity. Camera: static, eye-level, slight tilt down. Lens: 50mm, shallow depth of field, bokeh in the background. Action: she pumps two drops of serum from the bottle, gently presses it into her cheeks, looks up at her reflection with a small satisfied smile. Lighting: soft morning light from window on her left. Pacing: deliberate, no quick movements."

The strong version produces a coherent, watchable clip in two of three attempts. The weak version produces something that looks vaguely like an ad in maybe one of five attempts. The difference is that the model now has enough specification to make consistent choices about everything you didn't explicitly mention.

The camera and lens specs in particular are doing more work than you might guess. "50mm shallow depth of field" tells the model to compress the background and create blur — that single phrase changes the perceived production value substantially.

Anti-patterns: the prompts that produce mediocre output

Three patterns repeatedly produce disappointing results and should be avoided:

The kitchen-sink prompt. "A young athletic woman with brown hair tied in a ponytail wearing pink athletic clothing running through a forest path with autumn leaves in the morning light with soft bokeh background and dramatic motion blur shot on a Canon EOS R5 with a 70-200mm lens in cinematic style with film grain and color grading inspired by Wes Anderson." Too many adjectives compete for the model's attention and you get muddled output. The model is now trying to satisfy fifteen constraints simultaneously and will half-satisfy most of them.

The mood-board prompt. "Like a Calvin Klein ad but with more energy, kind of Apple-meets-Nike vibes, sort of Y2K but modern." Models do not have crisp shared vocabularies for cross-brand aesthetic mashups. Translate the vibe into specific visual elements: lighting type, color palette, composition style, subject framing. Then prompt those elements directly.

The negation-without-replacement prompt. "Don't make it look like a stock photo." The model has no idea what your specific "stock photo" complaint is. You need to say what you want instead, not what you don't want. "Use natural candid framing, slight off-center composition, ambient unstaged lighting, real-world textures" replaces the negation with positive guidance.

Iteration is part of prompt-writing, not failure

One of the highest-leverage realizations: the first generation is rarely the final one. Expect to iterate on the prompt itself two or three times based on what comes back. The prompt is not a magic incantation. It is a starting hypothesis that gets refined based on what the model actually produces.

The iteration loop looks like: write the prompt, generate four variants, examine what came back, identify the gap between intent and output, adjust the slot that produced the gap, generate again. Usually after two adjustment rounds the prompt is sharp enough that subsequent generations are reliably good.

The teams that get the most out of platforms like ours treat prompt-writing as a craft that improves with practice. The teams that struggle treat each prompt as a one-shot and blame the model when the first attempt doesn't land. Same tool, very different outcomes.

Building your team's prompt library

Every prompt that produces a winning ad should be saved to a shared library. Tag each entry with the ad's performance data, the product type, the placement, and any reference images used. After six months, this library becomes your team's institutional memory for what works in your category.

The library is also the fastest onboarding tool for new team members. Hand them the library on day one and they get a year of prompt-writing experience in a week of reading. Without the library, every new hire reinvents the same prompts and makes the same beginner mistakes.

Practical structure: a single shared document or spreadsheet with one row per prompt. Columns for the prompt text, the resulting ad ID, the platform it ran on, the performance numbers, and any notes on what made it work. Boring tooling. Massive compounding value.

The mental model that ties it all together

The right mental model for prompt writing is not "asking the model nicely." It is "writing a brief for a very fast designer who has never met you, knows nothing about your brand, and will produce something literally based on what you wrote." Every slot you leave blank, that designer fills with whatever was most common in their training. Every slot you fill clearly, that designer respects.

Once you adopt this mental model, prompt writing stops feeling mystical. It becomes a structured exercise in specifying enough that the model has a clear job, and not so much that the model is trying to satisfy contradictory constraints. That balance is the craft. With a few weeks of deliberate practice using the four patterns above, you can get reliably good at it. The reliability is the whole game.

One closing thought: prompt patterns will change as the models change. The five-slot pattern works well today because of how current diffusion models weight inputs. Six months from now, with new model architectures, the optimal patterns may shift. What will not change is the underlying discipline — write structured prompts, expect to iterate, keep a library of what worked. Those habits compound across every model generation.