How to write a product proposal with an AI coding agent
Work through your app idea with an AI coding agent, from intended users and screen mockups to platforms, technology, architecture, security and testing. Use a guided interview to record decisions and questions before implementation.
Whether you are developing an application, a kitchen appliance or, heck, a better screwdriver, you need to work out who it is for and why someone would want it. Product managers have long used proposals and briefs to turn those questions into something people can discuss before committing to the build.
When you are building your own software with AI coding agents, you take on several of those roles yourself. You have the idea, make the product decisions and judge the result, while agents help with the engineering. That makes pursuing a product of your own more practical, including ideas you might previously have set aside because you could not write all the code yourself. It also means the product manager’s work lands with you, whether or not you use that title.
A product proposal helps you do that work. You can develop it with your coding agent: describe the idea, let it ask questions, consider the options and record what you decide. You do not need a finished specification to begin. The aim is a short, useful account of who the product is for, what its first version should do and which questions still need an answer.
Giljo HQ includes three ready-made prompts to help you get started with your connected AI coding agent:
- Start from an existing codebase. Ask your agent to review the repository and use what is already built to prepare your product information.
- Bring a product proposal or vision document. Give your agent the document as the starting point for defining your product.
- Develop an idea through questions. Let your agent guide you through a conversation about what you want to build and record the decisions you make.
This article follows the third option: turning an idea into a product proposal before coding begins.
Let the agent interview you, one question at a time
Start with what you already know: your idea, intended users, existing code, preferred tools or constraints. Ask the agent to interview you before implementation. The result should cover the product, its technology and architecture, and how you will check the work.
Use the interview to get help where you need it. If you already understand the product and its users, ask the agent to spend more time proposing the technology and explaining how the parts should fit together. If you are comfortable with the engineering, ask it to challenge the intended workflow and help you explore the user experience. A rough screen mockup can make a vague idea easier to discuss. You can ask for alternatives, question a recommendation or say “I’m not sure” without bringing the conversation to a halt.
Here is a lightly edited version of Giljo HQ’s product-shaping prompt. It follows the four sections of the product card. You can use the same interview without Giljo HQ by asking for a Markdown document to keep with your project instead.
“Act as my product-shaping partner. Interview me ONE question at a time, never a wall of questions. Walk these areas in order (they mirror how Giljo HQ's product card is organized):
1. Product info: what I am building, for whom, and what "done" looks like (a name, a description, the core features).
2. Tech stack: programming languages, frontend frameworks, backend frameworks, databases and storage, infrastructure, and target platforms.
3. Architecture: the overall pattern, key design patterns, API style, coding conventions worth locking down, and brand/design guidelines if any.
4. Quality and testing: quality standards, a testing strategy (TDD, BDD, Integration-First, E2E-First, Manual, or Hybrid), testing frameworks, and a test-coverage target.
At EVERY question offer me explicit escape hatches: "I'm not sure", "can you guide me?", and, for tech-stack and architecture choices, "would you like me to propose one?". If I hesitate, propose a sensible default and justify it in one sentence, then move on.
When we are done, write a complete vision document in markdown, organized with those same numbered sections IN THAT ORDER (the card fills section by section from it), that I can upload to my Giljo HQ workspace (https://app.giljo.ai). End with a one-paragraph summary of what we decided.”
You do not need to know every term in that prompt. Ask the agent to explain unfamiliar choices through their consequences for your app. Treat its defaults as recommendations, not decisions you have silently accepted. Pause over choices involving sensitive data, ongoing costs or difficult migrations, and keep unanswered questions visible in the document.
Define the product and try the experience
For an illustrative example, imagine a hiking app that helps people find and compare trails. The interview should establish the main use: planning a walk at home, finding a route while outdoors, or both. That answer affects more than the feature list.
For a planning-only first version, you might agree on browsing trails, searching by name and comparing distance and difficulty. Live navigation, accounts and user-submitted routes could wait. Identify the source of the trail information, your permission to use it and what the difficulty ratings mean.
Ask the agent to sketch the main screens or make a small clickable mockup using sample data. Try finding a trail and comparing it with another. Review the layout, wording, navigation and accessibility before committing to the interface. The mockup demonstrates the proposed experience, not a working backend.
Choose platforms and a proposed tech stack
The tech stack is the set of languages, frameworks, storage and services used to build and run the app. Start with how it will be used: in a browser, on a phone, or as an installed Windows, Linux or macOS application. Decide whether it runs locally, is self-hosted or is a hosted service. A browser interface does not necessarily mean the data has to leave the user's computer.
For the hiking app, ask whether it must work on a trail without a mobile signal. An offline requirement changes how maps and trail information are stored and updated. Agreeing on that early is more useful than choosing a fashionable framework first.
Tell the agent about languages you know, existing code, your hosting budget and who will maintain the app. Ask it to recommend a suitable approach and explain the trade-offs. For storage, that might mean comparing a static trail dataset, a local SQLite database or a hosted database such as PostgreSQL. Those are options to assess against the requirements, not interchangeable defaults or a required progression.
Estimate how many people will use the app and how many might use it at once. Ask the agent what that means for the proposed design, and record any assumptions that need testing.
Plan the architecture and security
Architecture describes how the parts fit together. Ask for a simple diagram showing the interface, any backend service, data storage and external services. Have the agent explain how a request moves through them and where data is read or changed. If it proposes an API, the interface between those parts, ask why one is needed.
For a read-only trail catalogue, have it compare serving a prepared dataset with running a backend and database. If you later add private saved walks or route uploads, the proposal needs to address accounts, permissions and data changes. The architecture should follow those needs rather than assume every app needs the same components.
Establish what information is private, who may access or change it, where credentials belong, and how data will be backed up and recovered. Ask the agent to identify risks and choices that need further review. An architecture diagram or an AI recommendation is not a security assessment.
Record the main decisions and why you made them, along with relevant coding conventions and design guidelines. You need enough direction to guide the build, not a design for every function before you have tried the first screen.
Agree how the work will be tested
Quality and testing belong in the proposal too. Ask the agent to recommend checks suited to the app and explain what each will establish. The prompt names approaches such as writing tests before implementation, checking components together, testing complete user journeys and reviewing behaviour manually. You can combine them; you do not need to pick an acronym without understanding it.
For the hiking app, initial checks might cover finding a trail by a partial name, handling no results, comparing consistent distance units and using the interface with a keyboard. If offline use is included, test it without a connection. If private accounts are included, test that one user cannot read another user's saved walks.
Ask which failures matter most and how the tests would catch them. A coverage percentage alone does not tell you whether the app works correctly.
Review and save the complete proposal
Ask the agent to assemble the document in the same four sections: product info, tech stack, architecture, and quality and testing. Each section should distinguish agreed decisions from proposed defaults, assumptions and open questions. Include the reasons for significant choices and references to any mockups you reviewed.
Read it as the proposed basis for the work. Does it describe the experience you want? Are the platform and storage choices justified? Have hosting costs, access to data and maintenance responsibilities been discussed? Are the checks specific enough to tell you whether the first version works?
Correct anything the conversation did not settle. A short technical experiment may be the next step when a choice is uncertain. You can later expand the proposal into a detailed product requirements document, or PRD, without pretending that every design decision is already final.
Keep the agreed document somewhere you and your agents can retrieve it, and update it when the direction changes. In Giljo HQ, a vision document can be a starting point for your product information. Your connected coding agent reads it and helps populate that background. The reasoning happens in your chosen AI tool; Giljo HQ stores the context and work records.
Use the agreed proposal to choose the first piece of work. It gives you a product direction, a proposed approach and checks for the result.
Was this useful?
A cookie remembers your mark. Feedback privacy.