𝗔𝗜 𝗪𝗿𝗶𝘁𝗲𝘀 𝗖𝗼𝗱𝗲. 𝗜𝘁 𝗖𝗮𝗻'𝘁 𝗗𝗲𝗳𝗶𝗻𝗲 "𝗗𝗼𝗻𝗲."

I built an AI slide deck generator. You type a topic and get a presentation.

The AI demo worked instantly. It wrote an outline. It made slides. It exported a file. On my screen, it looked finished.

It was not finished.

The AI proved one user could make one deck one time. A real product is different. It needs to handle 100 users at once. It must bill them correctly. It must recover when a step fails. It must export a file that actually opens in PowerPoint.

The AI did not add these things because I did not tell it to.

When you build with AI, the hard part is not describing the feature. The hard part is defining what "done" means.

I used Velobase Harness for the foundation. It handled auth, payments, credits, and databases. This let me focus the AI on the PPT generation itself.

That is where "looks done" and "is done" pulled apart. I found four main gaps:

  • Concurrency: AI thinks one successful run equals a finished feature. A real system needs to split tasks into queues. You must generate each slide as its own job so workers can scale.

  • Billing: AI will deduct credits once and stop. A real product needs a state machine. You must reserve credits, settle based on actual usage, and refund on failure.

  • Self-review: If the backend retries a task silently, the user sees a spinning icon. You must show the status. The user should see "checking" or "redrawing" so they know the system is working.

  • Export: A slide can look great in a browser but break in a PPTX file. The requirement is not just to make a file. The requirement is that the file must match the web preview.

I changed my instructions. I stopped giving the AI a feature list. I started giving it acceptance criteria.

My new rules for the AI:

  • Build a production-ready SaaS, not a demo.
  • Support 100 concurrent users.
  • Use independent queues for planning and slide creation.
  • Connect every model call to the billing system.
  • Pause and resume tasks based on credit balances.
  • Run layout checks after every slide.
  • Ensure the PPTX export matches the visual preview.
  • Write tests for concurrency and failure, not just success.

AI writes code fast. It just does not know what makes code shippable. It treats a local demo as a full system.

인간의 역할은 엔지니어링 경계, 실패 사례, 그리고 비즈니스 규칙을 제공하는 것입니다.

Harness는 SaaS 기반을 제공합니다. AI가 로직을 채웁니다. 여러분은 그저 가장 정교한 입력을 제공하기만 하면 됩니다.

만약 AI로 구축한 제품을 출시했다면, "작동한다"와 "고객에게 선보일 준비가 되었다" 사이의 가장 큰 격차는 무엇이었나요?

출처: https://dev.to/velobasex/ai-can-write-the-code-it-cant-tell-you-when-the-product-is-done-4oh6