Work with an AI cofounder from an idea to a working pilot

Use AI to challenge an idea, define a small pilot, create the repository, and connect Railway and Supabase without confusing output with validation.

An AI cofounder is most useful when it helps you make better decisions as well as produce work. It can challenge a vague idea, turn assumptions into questions, and help build a pilot you can put in front of someone.

You still own the decisions, accounts, and commitments. Treat the agent as a capable collaborator whose work needs evidence, rather than a substitute for customer contact.

Begin with the assumption that matters most

Suppose you want to create a customer onboarding planner for independent consultants. Before asking for screens, ask why the current process fails. Do consultants need a planner, clearer scope, or help getting clients to respond?

Have the AI list those alternatives and explain what evidence would distinguish them. Talk to potential users before turning one explanation into an architecture.

Give the agent a role and a finish line

Use ChatGPT Work or Codex with Astra Extra High where available:

Act as my product and technical collaborator for a small onboarding-planner pilot. Challenge the problem definition, identify assumptions, and propose the smallest workflow that can test the idea. Then create an authorized private GitHub repository, prepare deployment to Railway, and use Supabase for private saved plans. Keep billing, automated client messages, and complex integrations out of version one. Finish with tested behavior and remaining uncertainties.

Set a time and cost boundary. Ask the agent to surface choices that would change the scope instead of quietly expanding the application.

Keep a decision record

Store a brief explanation of who the tool serves, the first workflow, excluded features, and the acceptance criteria. Record why important choices were made. This helps a later AI session, contractor, or collaborator work from the same understanding.

The repository should contain the source and setup instructions. Railway should run a known branch and commit. Supabase should have a reproducible database setup and explicit access rules. Keep credentials out of the repository.

Test the first promise

For the planner, a user should create a plan, edit an activity, return later, and see the correct saved version. Test the deployed URL and a second account's access. Include clear handling for a failed save.

Then ask a potential user to complete the task while you observe. Does the structure match how they think? Does it help them do something they already needed to do? What do they still handle elsewhere?

Let the collaborator challenge the result

Bring the observations back to the agent and ask for competing explanations. A confusing label, an unnecessary workflow, and a weak problem definition require different responses.

Review these pilot observations. Separate usability problems from evidence that the problem or audience is wrong. Recommend one next experiment and explain what result would cause us to change direction.

Explore Railway through our referral link when the pilot is ready for hosting. A useful AI cofounder workflow produces both a working artifact and a more honest understanding of the opportunity.

Continue with Explore the AI Cofounder journal.

Sources

• OpenAI: Get started with ChatGPT Work
• Railway: GitHub autodeploys
• Supabase: Row Level Security
• Supabase: API keys
• Railway: Affiliate program