
These are SaaS product videos, and every startup wants one because they are smooth, premium, and full of motion design. Making one takes a skilled editor days of work in After Effects, and right now the internet is full of videos saying AI can do this for you with just one prompt. However, most of those are simple demos, and almost no one has tested it on a real client project. We got a project from a client called VoiceWala to make a SaaS launch video for their AI app. The challenge is that I am not allowed to open my editing software, Premiere or After Effects; AI has to make every visual while my only job is to direct. I want to find out if AI can make something worth the client’s budget, show how a real product works screen by screen, and come anywhere close to my senior editor’s version. Before we start, please hit the subscribe button if you enjoy our content, as it helps us create better videos for you.
To get started, I copy-pasted everything the client sent me into Claude and asked it to create an instruction file containing my identity, the client’s style direction, references, and folder assets. Claude broke the process into phases—prep, style frames, voice and audio, animations, self-check, and revisions—so we wouldn’t waste tokens or credits. VoiceWala is an AI voice platform for Indian languages with over two hundred voices and different emotions, similar to an Indian version of ElevenLabs, and they wanted a product launch video. The client only provided the voiceover, script, website assets, and a single screenshot of the platform. I added rules to keep the UI real, follow brand colors, hide pricing, and ask for permission before spending any credits.
Once the session started, Claude struggled at first because it couldn’t watch video references directly, though it eventually reviewed them. It transcribed the voiceover word-by-word so on-screen text would land right on cue, and then rebuilt the single app screenshot from scratch in code while designing matching screens for better animation control. The first version was ready as a basic explainer video. While it stuck to the brief, it lacked the dynamic SaaS flow we needed. For the next version, I asked it to build a 3D carousel of sixteen voice cards that whips around to land on each speaker’s first syllable—a task that would require tedious frame-by-frame work in After Effects—and it managed to sync the animation to the voiceover.
Because the tutorial section still felt too static, I grabbed Behance inspiration featuring laptops and phones floating in nature with a liquid glass aesthetic. Since I couldn’t use After Effects mockup templates, I connected Higgsfield MCP to Claude to generate the device scenes. Claude then wrote its own Python scripts to track the four corners of the green screen on every frame, pin the coded UI on top, key out the green, and even suppress green spill reflections on the keyboard. However, that version turned out terrible because clips repeated endlessly, the pacing was slow, and it wasted hundreds of generations because Claude only checked thumbnails instead of watching the full video.
Frustrated and worried about comparing this to my team’s work, I set strict new rules: Claude had to act like a professional animator, get my approval on a written storyboard before building, and export every render frame-by-frame to watch as a moving video before showing me. This helped it catch mistakes and build a new plan, though its music and sound design mix was still awful. Since a video relies on audio flow, I decided to use my team’s completed audio track—which also ensured a fair side-by-side visual comparison at the end. Claude impressively used Meta’s Demucs AI model to separate the track into voice, music, and nearly sixty individual sound effects, mapping the rhythm so every visual cut landed on a beat or whoosh.
Tired of typing long feedback notes, I asked Claude to build a Frame.io-style review dashboard, and it created a working video player artifact where my team and I could leave comments. While the new outdoor shots looked better, I also used today’s sponsor, Storyblocks—a completely human-made stock media library with unlimited downloads, commercial licensing, and a new MCP integration that lets you search and download assets directly inside Claude or ChatGPT. Through multiple rounds of feedback—even installing Wispr Flow to dictate my notes by voice—we went through weird particle intros, broken laptop physics, a space-theme experiment, and a light-and-dark theme switch across multiple drafts. Finally, I shared my team’s video purely for structural inspiration, added a Dropbox-style 3D coded phone intro, and had Claude scan frame-by-frame to fix thirty-three sudden jolts and three frozen moments.
In the end, Claude’s version went through seven versions and thirty revisions, while my team built theirs from scratch alongside it. When we played the final VoiceWala launch videos side by side and scored them, my team’s edit scored higher overall, beating Claude’s version. My team scored top marks in typography and motion transitions for creating a cohesive piece, while Claude scored much lower due to repetitive presets and choppy start-and-stop pacing, though both tied for clearly explaining the product. Ultimately, this was a heavy project built completely from scratch that normally requires a very experienced editor. While AI lacks human creative taste and cannot yet handle massive amounts of raw footage, it is a great tool for reducing repetitive tasks, and I strongly recommend editors and motion designers start incorporating it into their workflows.