---
description: Why a working demo is only the beginning: goals, error cases and traceable checking steps.
url: https://www.schmalzl.tirol/en/blog/2025-10-04-ki-prototyp-verlaessliche-anwendung
lang: en
date: 2025-10-04
deutsch: /blog/2025-10-04-ki-prototyp-verlaessliche-anwendung.md
---

# From AI prototype to a reliable small application

_4 October 2025 · Archive date_

A small application is quickly demonstrated: type in three entries, click save, done. The next question is harder: what happens if the window is closed, a field is left empty or an entry appears twice?

This is where the path from prototype to reliable tool begins. Even if an AI helps with the writing, the result needs a clear idea of what is supposed to work.

## An example with an invented book list

The task sounds simple at first: add books and find them again. For a usable first version, I would set down four properties: a title must not be empty, saved entries survive a reload, a search finds matching titles, and accidental deletion can be prevented or undone.

These are verifiable statements. "Make a nice book app", on the other hand, is an invitation to many decisions whose result is hard to judge later. A good GUI can hide the fact that saving only works while the window is open.

## Small changes need an end point

A change should solve one concrete task. Then come the matching checks: try the normal flow, provoke an expected error and check whether existing functions still work. Only then comes the next extension.

The publicly documented [Superpowers workflow](https://github.com/obra/superpowers) is a find related to this step-by-step way of working. It is no guarantee of quality. What matters is the result in the project.

A first version does not have to do everything. It should do a few things understandably and reliably. That is less spectacular than a long feature list, but much more pleasant as soon as the tool is actually used.

## Original finds

These links were the starting point for this post. The commentary in the text is mine; the linked posts are not my own tests.

- [What lies between a vibe coding demo and a product](<https://www.linkedin.com/posts/anthony-sidashin_most-people-today-use-ai-for-emails-linkedin-activity-7378139842788106240-33Gm>)
- [What lies between a vibe coding demo and a product](<https://www.linkedin.com/posts/alexander-schikowsky-2a01781b3_du-kannst-software-an-einem-wochenende-mit-share-7450796272015335424-e4OC>)
- [What lies between a vibe coding demo and a product](<https://www.instagram.com/p/DXwGePYDawL/>)
- [Coding agents: first the task, then the tools](<https://www.linkedin.com/posts/juliadanyal_everyone-wants-to-build-ai-agents-almost-share-7464340186819162112-6CQn/>)
- [Coding agents: first the task, then the tools](<https://www.linkedin.com/posts/shawn-wolter-93a263195_why-karpathys-claudemd-got-48k-stars-share-7464672827250683904-lQGR/>)
- [Recurring review loops in AI coding](<https://www.facebook.com/share/18XDtK53FA/>)
- [Own the Outer Loop: responsibility in AI development](<https://www.linkedin.com/posts/marcel-more_own-the-outer-loop-activity-7483739497566928897-4kZr>)
- [Superpowers as a structured development workflow for coding agents; check the claims about success and reach.](<https://www.linkedin.com/posts/a-developer-on-github-just-built-a-full-development-share-7475814139605577728-BNQT/>)

> **About the date:** This post was written on 11 October 2026. The archive date corresponds to the earliest related find in the collection, not to a publication at that time. Technical context and sources added later refer to the time of writing.