How to use Gemini to build a verified voice intake agent
The InputRight founder used an architecture LLM, an engineer LLM and Gemini CLI to build an open-source voice intake MVP. Users verify transcribed data in a form by confirming or editing it.
news.ycombinator.com
Prompt to try
Editorial adaptation, not the source author’s exact prompt. Replace bracketed inputs and review tool permissions first.
Outline [voice intake project] for [required fields]. Produce commands and code for implementation with Gemini CLI. Analyze [command outputs and errors] and give the next step. Display transcribed fields for user confirmation or editing by voice or manually. Prepare a continuation prompt when needed.
Reach people exploring AI agents
See placements →What happened
The author reports a working open-source MVP on GitHub in less than four weeks. Its voice transcription is displayed for user confirmation or editing, which the author claims ensures complete data accuracy.
Reported by Jeff_Morton_AI on Hacker NewsHow the agent did it
- Ask a custom-prompted LLM to outline the project architecture.
- Pass the architecture to an engineer LLM for commands and code.
- Run the supplied commands through Gemini CLI.
Where a person stepped in
The author copied outputs and errors between models. In the resulting product, users confirm or edit their transcribed input before it is accepted.
A person guided the agent closely as it worked.
Jeff, the founder of InputRight, wanted a website voice agent for a contractor lead-generation service. His concern was that a misheard phone number could undermine a valuable lead. Without a development background, he used a top-down LLM workflow to turn the idea into a working open-source MVP in less than four weeks, according to his post.
First, a custom-prompted LLM in Google AI Studio acted as a CTO and outlined the project structure. The author passed that plan to another LLM acting as a senior software engineer, which supplied commands and code snippets. Gemini CLI then acted as the junior developer running those commands. The author copied outputs and errors back to the engineer model for analysis and instructions about the next step.
Debugging remained difficult because the author depended on the LLM to interpret errors. He also reports declining performance when Gemini 2.5 Pro context reached roughly 600,000 to 700,000 tokens. He addressed this by asking the model to generate a continuation prompt for a new AI engineer.
InputRight transcribes a user's speech, displays the result in a form, and asks the user to confirm or edit it by voice or manually. The author claims complete data accuracy through that verification step; the post supplies no independent accuracy test.
Who it's for: Builders of voice intake tools who want users to verify captured details before accepting them.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Hacker News · Jeff_Morton_AI
Frequently asked questions
How were the AI development roles divided?
A custom-prompted LLM in Google AI Studio outlined the architecture, another LLM supplied commands and code, and Gemini CLI executed commands. The author relayed outputs and errors back for advice.
How does InputRight verify captured data?
It transcribes speech, displays the transcription in a form, and asks the user to confirm or edit it through voice or manual input. The author claims this produces 100% accuracy; no independent test is provided.
How did the author handle growing context?
After noticing performance degradation around 600,000 to 700,000 tokens in Gemini 2.5 Pro context, the author asked for a prompt that would let a new AI engineer continue the project.
More details
Remaining steps
- Relay command outputs and errors to the engineer LLM.
- Generate a continuation prompt when the context degrades.
About this account
LLMs outlined the architecture, generated code and commands, and analyzed errors; Gemini CLI ran commands within a human-mediated workflow.
Platform: Hacker News. Attribution: https://news.ycombinator.com/item?id=44938322. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: Jeff_Morton_AI on Hacker News
Source published 18 Aug 2025
Source checked 14 Oct 2026
Tools and topics
Google AI Studio · Gemini CLI · Gemini 2.5 Pro
Voice intake · Human verification · MVP
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.