Inspiration
The inspiration is one of those internet jokes that never really dies. Someone films a friend doing something completely mundane, like digging through the fridge or losing a fight with a vending machine, and dubs it with a hushed, serious nature documentary voice. It's been a meme format for years because it's funny every time, but making one yourself is a pain: you have to write the narration, find a voice, and sync it up manually. We wanted to take a joke people already loved and remove all the friction between "having the idea" and "having the clip."
What it does
Attenborough Mode lets you either record yourself (or a friend, ideally without them noticing) live in the app, or upload an existing 15–30 second video. The app then generates a fully narrated nature-documentary voice-over based on exactly what's happening in that specific clip — not a generic script, but one that reacts to the actual moment you captured and lays the narration directly on top of the video, ready to share.
How we built it
The whole pipeline runs on Gemini. While recording, we use Gemini's live camera tracking to follow what's actually happening in frame in real time. That description gets passed to the Gemini API along with a tightly constrained system prompt. Reverent, hushed, deadpan British naturalist tone, no slapstick, no fourth-wall breaks, no real names, roughly 60–90 words which writes the narration script for that specific clip. Finally, we run that script through Gemini's voice AI/text-to-speech to generate the narrator audio and mux it onto the original video.
Challenges we ran into
It's easy for an LLM to drift into either flat description or actual jokes/punchlines, and the whole bit only works if it stays deadpan and reverent about something completely ordinary. We spent a lot of time tuning the prompt constraints (word count, banned phrasing, no addressing the viewer) to keep the voice reliable across very different kinds of footage. Syncing generated narration timing to variable-length clips, and making the live camera-to-script-to-voice pipeline feel fast enough to not kill the joke, were the other big technical hurdles.
Accomplishments that we're proud of
We're proud that the narration genuinely feels reactive. It's not a template with mad-libs blanks, it's actually describing your specific clip beat by beat, which means it holds up as funny even after you've used the app a dozen times. We're also proud of how tight the end-to-end pipeline is: record or upload, and in about thirty seconds you have a finished, shareable clip.
What we learned
We learned a lot about prompt-constraining a model to hold a very specific comedic voice consistently, rather than just getting "a funny response" — reverence and deadpan restraint turned out to be much harder to enforce than jokes are. We also learned how much of the product experience lives in orchestration: none of the individual pieces (vision, scripting, TTS) are exotic on their own, but chaining them together smoothly was most of the actual work.
What's next for Attenborough Mode
Next up: additional narrator voice/style options, a consent-check flow for clips involving other people before they're shared publicly, batch processing so users can generate multiple clips at once, and eventually opening up the underlying "watch it, write about it, say it" pipeline as an API other creator apps could plug into.
Log in or sign up for Devpost to join the conversation.