Watch on YouTubeJoe Reeve built an app that identifies a statue, researches its history, designs a matching voice, and starts an ElevenLabs agent conversation in about 30 seconds.
The prototype took two hours to build in Cursor, then reached 50,000 impressions before a follow-up post about vibe coding reached 1.5 million.
Voice agents work better when audio is paired with visual information, explicit interaction cues, and ways to interrupt or skim a response.
Joe Reeve explains how he built a statue app in two hours on a Sunday with Cursor and a single prompt. A phone photo goes to OpenAI deep research, which identifies the object and produces historical material plus a proposed voice description. ElevenLabs' voice design API creates the voice, then an ElevenLabs agent starts a phone conversation. The pipeline takes about 30 seconds. The project spread from 50,000 impressions to 1.5 million after Reeve posted about vibe coding, leading museums, auction houses, and travel companies to ask about similar systems. Reeve is direct about what remains unfinished. Museums need curators to write the narrative and connect the app to their own collection databases. He then discusses voice interfaces, arguing that people need visual output, better interruption cues, and ways to skim spoken answers. He also describes possible consumer applications for vibe coding, including social games built around simple shared primitives.
The app starts when someone photographs a statue. OpenAI deep research identifies the object and generates historical knowledge, along with a description of what the statue's voice might sound like. ElevenLabs' voice design API turns that description into a voice. The app then creates an ElevenLabs agent and starts a phone call. Reeve says the complete sequence works in about 30 seconds, from taking the picture to speaking with the statue.
Reeve built the app in Cursor on a Sunday because he was tired and bored. He published the prompt he used to build it in one shot. The initial post and video received 50,000 impressions. After he reposted the project with a comment about vibe coding, it reached 1.5 million impressions the next day. Museums, auction houses including Bonhams and Christie's, and businesses similar to Tripadvisor then contacted him about letting people talk to objects in their collections.
Reeve says the app is mostly a collection of APIs designed to scale. User management could be added with a service such as Supabase, while agent operation and voice design would remain with ElevenLabs' APIs. He thinks vibe coding changes where project value sits. Connecting services and describing the interaction clearly can matter as much as solving a difficult technical problem. He still says the prototype needs more work before it is production-ready.
Reeve does not see a photograph followed by automatically gathered research as the long-term museum experience. Curators need to decide the actual narrative instead of relying on random material found through search. He notes that museums often treat their collection databases as core intellectual property, and many have APIs that could provide better source material. His current idea for curator editing is a dashboard where people change the system prompt and knowledge files, though he expects text editing to be the main interaction.
Reeve describes conversations with Jago, formerly head of the Americas at the British Museum and now running the Sainsbury Centre. They considered how an inanimate object should sound when its material came from one place, was carved in another, and spent centuries in a British museum among visitors from many countries. The answer could involve several cultural and linguistic influences. Reeve thinks this question will matter as companies give voices to objects and everyday systems such as lifts.
Reeve says voice interfaces have a basic limitation because users usually have to choose between voice and another interface. He imagines speaking to a product-manager agent that passes work to a coding agent, rather than talking directly to the system doing the work. He also expects more multimodal conversations, where the agent's extracted information appears visually while the user speaks. People hesitate to interrupt agents, so products need to make interruption feel acceptable and give clearer signals about when the agent wants to respond.
The discussion turns to information density. Voice is useful for getting thoughts out quickly, but people may want diagrams, text, or a generated interface in response. Reeve says he wants single-input voice with parallel output in the form of the application or visual material. The group considers controls for moving forward and backward through audio, perhaps by concepts rather than sentences. They also discuss showing summaries of sections that users can expand, similar to interaction patterns in the Claude app.
Reeve thinks consumer vibe coding needs a simple social format. He points to Facebook Instant Games, which offered primitives such as user and friend information, leaderboards, basic data storage, and rewarded or interstitial ads. He bought a Fruit Ninja clone, added the old API, and woke up to 15 million users, mostly in Vietnam, although he made little money. He sees that combination of simple creation, social sharing, and built-in distribution as a possible template for consumer-facing vibe coding.
Reeve made the statue video on his phone with CapCut and a borrowed DJI Bluetooth lapel microphone. Editing took about 20 to 25 minutes. He says his videos usually have a median view time of six to twelve seconds, so the opening hook needs to arrive quickly. Captions help, and music changes the feeling of a video even when it is quiet. He sometimes chooses the music after making the narrative, and sometimes starts with a musical mood and writes the speech around it.
"So you take a picture of something you get all the search research back from OpenAI generate a voice and start talking to an agent to a statue within 30 seconds."03:06