Why We Have a Prompt Library Instead of Improvising for Every Campaign
A good prompt is almost never born on the first try. It takes several iterations to discover which phrasing gets what you want out of the model, and without documenting it anywhere, every new campaign starts from scratch again. That’s why we built an internal prompt library: about 20 recurring templates across video, image, and audio models, each with notes on when it works and when it doesn’t.
This isn’t a “magic prompt” that solves everything. It is a list of starting points that saves the repeated cycle of trial and error, allowing the team to start at 80% ready for every new incoming brief.
The library has a less obvious benefit: a new team member starts with templates already tested across dozens of real campaigns, instead of learning through trial and error what works and what doesn’t. They focus on adapting to the specific brief, not building a foundation from scratch.
It also helps in conversations with clients. When a client asks why we chose a specific camera angle or scene description, we can show them the documented template and the rationale behind it, rather than a generic answer like “this is what usually works, trust us.”
Video Prompts: Veo, Kling, Seedance
In video, most of our prompts are built of three parts: scene description, camera movement, and lighting. A few recurring templates:
- Product scene: slow zoom on a product on a clean workspace, soft side lighting, no cluttered background.
- People movement: natural walking toward the camera in an open space, daylight, without exaggerated hand gestures that the model tends to distort.
- Scene transition: one continuous camera movement between two focal points, instead of two separate cuts that are difficult to connect in editing.
- Brand atmosphere: close-up on texture (fabric, wood, metal) with dramatic lighting, to be used as a background for separately overlaid text.
In each of them, an explicit instruction appears at the end not to render text or letters in the scene, so the model doesn’t try to “help” and create distorted text instead.
Each template also has a documented “failure version”: a brief description of what happens when phrasing it slightly differently, less precisely. This helps new team members understand why this specific phrasing was chosen, rather than one that sounds almost identical.
Image Prompts: Nano-Banana Pro and Gemini 3 Pro Image
In images, the distinction is between creating from scratch and editing existing material. Recurring templates we use:
- Clean product background: product on a neutral surface, studio lighting, three-quarter angle.
- Background editing: replacing the background in an existing image without touching the product itself, preserving shadows and lighting.
- Color variation: the same composition in several brand shades, for quick A/B testing between versions.
- Lifestyle atmosphere: a natural product-in-use scene without identifiable figures, for general use across multiple campaigns.
Here too, the “no text, no letters” rule appears in every prompt we output, even when the template seems simple.
Another template we haven’t mentioned: “product-in-hand”, meaning a hand holding the product at a natural angle. Useful for ads that want to feel personal and less “catalog-like”. Natural lighting and a clean background yield the most credible result there, and the fingers are the first element worth checking before approving and sending to the client.
Audio Prompts: ElevenLabs and Stable Audio 3
In audio, the prompt describes mood and pacing, not exact words. A few templates:
- Short energetic background, without words, for the fast editing pace of a performance ad.
- Calm background for a long brand video, with a gradual build toward the CTA.
- Voiceover in a warm-neutral register, medium speaking pace, without unnecessary drama.
Audio templates depend heavily on context. A template that works great for a product video can sound out of place in an emotional brand video, which is why each template is also documented with the type of campaign in which it was tested.
We also specify the duration tested for each template. Music that sounds great for ten seconds can get tiresome when stretched to thirty seconds without change, and the reverse is also true.
The Rule That Repeats in Every Prompt in the Library
In every template—in video, image, and audio—there is a standard line: the model does not generate text that remains on screen. This isn’t a lack of trust in the models. Their job is to produce a scene, not to act as font software.
This line is present even in templates where there is no intention of adding text. The routine prevents mistakes, because you never have to remember to check for it separately. Whoever builds a new template copies this line first, even before writing the scene description.
How the Library Is Maintained and Updated
Each template is documented with the full original prompt, the failure note if there was one, and the fix that resolved it. It also includes the last test date, so it’s clear which templates haven’t been tested against the models running today. When a new model enters production, we check which templates are still relevant and which need to be rewritten, rather than assuming that what worked in the previous model will work just as well in the new one.
The testing takes a day or two with every new model rollout, and sometimes reveals that a template that worked great before now produces a different result, for better or worse. Without it, it’s easy to assume the template “still works” and discover the issue only after it has already been produced for a client.
Have a creative concept you want to test against our library? Send us the exact brief and we’ll tell you if there’s already a suitable template, or if we need to build a new one specifically for it.