The Character Sheet Was the Unlock. How I Fixed AI Video's Biggest Problem and Made a 35-Second UGC Ad.

The cinematic ad was about proving the concept. The UGC ad was about proving something harder: could AI generate a believable human, talking naturally to camera, with consistent character and working lip sync, across a 35-second ad?

The short answer is mostly yes. But the thing that made it work was not the generation tool. It was a single document I created before touching any tool at all.

The Problem With AI Video Nobody Talks About Enough

Character drift. Frame one and frame fifteen look like two different people. The hair is slightly different. The lighting shifts. The facial structure drifts. For atmospheric product shots it is forgivable. For a UGC ad built around a single believable human, it breaks everything.

This was the primary failure point in my previous attempts at character-led video. The cinematic TEEKHA ad avoided the problem by keeping the human presence minimal. This time the entire ad depended on one person staying visually and emotionally consistent across 35 seconds.

The fix was a character sheet.

What a Character Sheet Actually Is in This Context

Not a design document. Not a mood board. A single reference block of precise descriptive text that goes into every generation prompt as a prefix.

For the TEEKHA UGC ad, the character sheet described: approximate age, skin tone, hair length and style, clothing down to the color and fabric weight, camera angle, lighting conditions, and the specific emotional register for each section of the video.

The difference between “person looks tired” and “person slumped slightly forward, eyes at half-mast, chin resting on one hand, looking directly into camera with flat expression” is the difference between Seedance guessing and Seedance following.

Every single generation prompt for this video started with that character description before any scene direction appeared. The result was that the person in the opening frame and the person in the closing frame looked like the same person. That had not happened in earlier attempts without it.

The Format and Why It Suited the Brand

TEEKHA is a fictional Indian chai brand I built as a testing subject earlier in this series. Saffron and forest green, a roaring tiger logo, and a product called Laal Mirchi Chai that contains actual chilli. The previous video was a cinematic 15-second product ad. This one was a UGC-style ad built around a relatable human moment: the 3pm slump, the desperate search for something that actually works, and the mildly suspicious reaction to being handed a chai that bites back.

UGC ads work because they feel like a recommendation from a real person rather than a brand. For TEEKHA specifically the product story was already perfect for this format: a spicy chai that genuinely surprises people is exactly the kind of discovery someone films themselves talking about.

The Full Workflow

Step 1: The storyboard. A full 45-second script broken into three 15-second chunks. Each chunk had its own storyboard with frame-level timestamps and specific direction notes. General descriptions do not give Seedance enough to work with. Precise ones do.

Step 2: The character sheet. Written before any generation began. One document, precise physical and emotional description, used as a prefix on every prompt across all three chunks. This is the step most AI video workflows skip and it is the step that changes the output most significantly.

Step 3: Generation in Magnific Spaces. Each 15-second chunk generated separately using Seedance 2. The character sheet prefix kept the visual identity consistent. The timestamped storyboard gave Seedance the pacing and dialogue structure to follow.

Step 4: Assembly in CapCut. All three clips stitched in sequence, color-graded slightly, and trimmed in post. The 45-second storyboard became a 35-second final cut after removing frames that did not land cleanly.

What Worked

Character consistency held across the full 35 seconds. That was the goal and it was achieved.

Lip sync was on point throughout. Dialogue delivery felt natural rather than mechanical in most sections. The dry humor in the script came through in the performance, which is not something I expected AI video to handle well at this stage. The 3pm slump energy, the suspicious look at the TEEKHA tin, the slightly-too-honest reaction to the chilli: all of it read correctly on screen.

What Still Needs Work

Text on screen is where AI video still shows its limitations most visibly. Any time text appeared in the generated frames it was garbled, inconsistent, or simply wrong. This is the same limitation that appeared in the ChatGPT Images 2.0 testing earlier in this series and it has not been resolved in video generation. Any text element in a generated video needs to be added manually in post.

A handful of transition frames between dialogue sections felt slightly mechanical, noticeable on close viewing but acceptable for a UGC format where rough edges are part of the aesthetic.

The Comparison With the Cinematic Ad

The cinematic ad was visually stronger in individual frames. Product shots, macro sequences, atmospheric lighting: these are where AI image-to-video currently performs best.

https://video.wixstatic.com/video/4b58a0_362423b94a48416ca6a7cad538910c11/720p/mp4/file.mp4

The UGC ad was more effective as a complete piece of communication. A believable person saying believable things about a product is more persuasive than beautiful product shots, which is exactly why UGC outperforms cinematic ads on most D2C platforms right now. The format suited both the product and the current capabilities of the tools better than the cinematic approach did.

The Repeatable Process

The character sheet plus timestamped storyboard plus chunked generation is the workflow that produced consistent results. It is more work upfront than a single storyboard prompt. The output justifies it.

For anyone running D2C brands or client campaigns: a first-draft UGC ad is now within reach without a production budget, provided the thinking happens before the generation does. The tools reward preparation. They do not replace it.

TEEKHA has two completely different ad formats now. The brand is still fictional. The ads are not.