Lavendr Coffee
When a brand needs content that feels human, the workflow behind it has to be built that way too.
Content Production
AI Workflow Design
Role
Senior Video Editor / AI Content Producer
Timeline
Ongoing
team
Solo
platform
Meta Paid Ads
The Real Problem
Lavendr is a purple ube collagen coffee brand that needed video content that actually looked real. Not polished in a corporate way, but real in the way organic content feels real. The kind of content where someone scrolling stops because it looks like something a person actually filmed and posted.
The challenge was that most AI-assisted video production at the time was producing content that felt obviously generated. Faces that shifted between clips, product colors that looked wrong, movements that felt stiff. The brand needed a full content pipeline that could produce multiple ad concepts consistently without losing that human, authentic feel.
The feedback from early rounds was pretty direct.
"It needs to look like a real person filmed this."
"The avatar keeps changing between clips."
"The product color looks wrong, it's not purple enough."
Nobody had mapped out a complete AI-assisted video workflow specifically for this kind of branded content before. Each tool was being used separately and nothing was connecting into a coherent production system.
Finding the Fix
I started by mapping the full content pipeline from avatar creation to final export. After testing a lot of failed clips, three problems kept showing up consistently.
No consistent avatar. Every new clip generated a slightly different face. It broke the illusion that these were real people with real routines.
Wrong product color. The purple drink kept generating as brown or too dark. It looked like regular coffee, not the Lavendr product.
Unnatural movement. Clips were either too static or too exaggerated. Nothing felt like a real person casually holding a drink or going about their morning.
So I focused on three fixes. A two-image method where Image 1 is always the face or avatar reference and Image 2 is always the scene or pose reference. A product color rule where every prompt specifically called out "lavender purple" with negative prompts blocking brown or dark liquid colors. And a movement language system where I described actions as if a real person was doing them naturally, not as instructions to a machine.
All of this came together into a custom prompt skill file I built to document the full workflow and keep it consistent across every new concept.
What Actually Happened
I built out the full ad concepts and tested each prompt against reference images before generating final clips. That's honestly where I caught most of the problems before they became wasted time.
For example, I had been prompting the product preparation clip without specifying that the mixing tool should stay its original color throughout the clip. The AI kept coating it in white foam mid-clip. Once I added that rule to the negative prompt it fixed immediately and stayed consistent.
The voiceover integration also evolved. My first attempt used a basic prompt and the voice felt too clean and over-polished. It didn't match the casual authentic energy of the visual clips. I adjusted it using the ElevenLabs V3 system I built out, adding emotion cues and pacing markers so the voice actually sounded like a real person talking about something they genuinely liked.
The first concept went to the client and they noticed the improvement right away, specifically the product color and the consistency across clips. From there I rolled out the same system across all the other ad concepts.
What Changed
Production time dropped significantly once the two-image method and prompt templates were locked in. Concepts that used to take multiple days of trial and error started coming together in a single focused session.
The client feedback shifted completely.
"This actually looks like a real person filmed it."
"The purple color is finally right."
"I can actually see this running as an ad."
And honestly the best sign that the system was working was that the brand kept coming back with new concepts to produce. That's the real metric for me.
What I Had to Work With
No live talent. Everything had to be produced through AI-assisted tools, which meant I couldn't direct a real person on set. I had to figure out how to get consistent faces, consistent lighting, and consistent product color across multiple clips using only prompts and reference images.
A limited but specific toolchain. The brand was working with NanaBanana for image generation, Kling for motion, Seedance for multi-scene sequences, and ElevenLabs V3 for voiceover. Each tool had its own rules and limitations I had to learn through testing.
Fast turnaround. This wasn't a slow creative process. The client needed multiple ad concepts produced quickly, which meant I had to build a system I could repeat, not just figure it out one clip at a time.
These constraints shaped everything. Instead of trying to make each clip perfect from scratch, I focused on building a repeatable two-image method and a prompt structure that could scale across different concepts and scripts.
What I'd Do Differently
I'd lock in the avatar library earlier in the process. A lot of time in the early concepts went toward generating new avatar references for every clip instead of building a consistent set from the start. Having that library ready from day one would have made the whole pipeline faster.
I'd also document the negative prompts more systematically from the beginning. A lot of the rules I ended up building into the workflow came from mistakes I made first. Some of those mistakes cost real time that could have been avoided with more structured testing upfront.
What I Learned
A repeatable system beats a perfect clip. The biggest unlock wasn't any single great prompt. It was building a workflow I could run consistently and improve over time without starting from zero every time.
Specificity is everything when working with AI tools. Vague inputs produce vague results. The more specific I got about color, movement, lighting, and what I didn't want in the output, the closer the results got to what I actually had in my head.
The human feel is in the small details. A slight natural camera drift, a hand stirring a drink slowly, condensation on the cup. None of those things are technically necessary but they're what makes the content actually feel real. That's where I spend the most attention now and it's what separates content that performs from content that just looks generated.
