How to actually ship a product with an AI agent

A common failure mode of agentic tools is treating them like a chat partner. They're not — they're an executor with a leash. The folks who ship a working business in an afternoon hold the leash a specific way. Here's what we've worked out.

1. Lead with the *who* and the *what*, not the *how*

Bad: "Build me a Next.js app with Tailwind and a Postgres database for booking yoga classes."

Good: "I run a yoga studio in Bursa. I want a site that lets people book classes, pay, and get reminders."

The agent picks the stack. You pick the business.

2. Trust the first cut

The agent will make decisions you wouldn't make. Domain names, color palettes, copy tone. The temptation is to redirect every time. Don't. Let it ship the first version, then edit. You'll get further faster.

3. Be specific about constraints, not implementations

Bad: "Use Polar Checkout."

Good: "Customers should be able to pay in TRY and get a receipt as a PDF."

Constraints translate. Implementations don't — they encode assumptions about today's tools.

4. Use workflows for the boring stuff

Webhooks. Cron jobs. "When someone submits this form, send you an email and add them to the database." That's a workflow. Set it up once, never think about it again.

5. Show, don't describe

Drop a screenshot in the chat. The agent reads it. "Make my page look like this" + image works much better than 200 words of layout description.

6. Ask the agent to *audit* its own work

"Review the project for any obvious issues" → the agent will surface things you wouldn't have noticed (missing meta tags, slow queries, accessibility gaps).

The shape of working with an AI agent isn't talk to it like a human. It's give it a leash and watch where it pulls. Get the leash length right and you ship in hours, not weeks.