AI-generated database schemas—the structures that define what information a database stores and how that information relates—can look convincing. If the data saves without problems and the code runs, does that mean the design is good? Here are five examples of AI-suggested designs I reviewed and changed while building a real service.
1. Creating a statistics table without asking if it is needed
My service needs the subscriber count, video count, and total views for YouTube channels. Since these numbers change over time, I initially considered storing them in a separate statistics table, with a new row for each date. The AI suggested the same kind of date-by-date structure.
But the first question should have been: “Why do we need to keep collecting this data?” I could retrieve the subscriber changes for specific dates from an external API. Since the service did not need to store that data separately for its current analysis, I removed the table. The design itself was not wrong; the problem was that the AI never asked whether the service needed it.
2. Putting a rule that does not exist in a comment
In the comment for the column that stores subscriber counts, the AI gave “subscriber-count condition for a publishing destination rule” as an example. But our service has no such rule. Subscriber counts are used to analyze performance by comparing channel size with video views.
A comment may do more than simply explain a field. Another AI might mistake it for an actual policy and generate code based on it. Descriptions of service policies or data usage should be checked carefully to make sure they are accurate.
3. Overlooking the order in which features are used
To create promotional posts, we need to know the original video's language. But some YouTube channels do not have a default language set. The AI suggested reusing language data from the multilingual subtitles feature.
The problem is that users do not necessarily use channel analysis, multilingual subtitles, and video promotion together. If someone uses only the promotion feature and has not used multilingual subtitles, the required language data may not exist. It is not enough to check whether reusable data is available; we also need to consider the order in which users might use the features.
4. Overlooking the project's naming and time conventions
The AI initially named a table t_creator_profile. That name suggests a broad creator profile, but the actual columns contained only YouTube channel information. So I changed the name to t_youtube_channel and t_youtube_channel_profile. A name may work for now, but if it does not match the meaning of the data, it can cause confusion later.
I also checked the time conventions. After I told the AI that the server time was UTC, it applied UTC to date columns as well. But timestamps and dates need to be treated separately. A timestamp may be stored in UTC, while a date shown to a user may need to follow that user’s time zone. I added this rule to CLAUDE.md.
The AI named the measurement timestamp measure_time, but this project uses create_time for the creation timestamp. If a name feels unfamiliar, check whether it matches existing conventions and whether the two timestamps mean the same thing.
5. Suggesting a solution that does not match the service's purpose
The goal was to automate not only the creation of promotional posts but also publishing them to multiple sites. For sites without an API, the AI suggested providing a draft and a copy button, then having a person publish it manually. That is technically possible, but users would still have to move between sites and paste the content themselves. It did not match the goal of automation.
Instead, I designed the flow so users could use their own AI for sites without an API. The server would provide a single file containing a draft and publishing instructions for each destination, and users would pass that file to their own AI.
The AI then proposed a structure in which it would report the publishing result, post URL, view count, and comment count back to the server. But continuously managing data from sites without APIs would require a separate collection system. In my view, that data went beyond what the service needed for its purpose.
Why you need to review the DB schema yourself
These five cases were not problems where the code failed to run. They were easy to miss precisely because everything worked. The fact that data can be saved does not, by itself, make a design good. I keep coming back to these questions:
“What happens if this value is missing?”
“Who fills in this value, and when?”
“Why do we need to keep collecting this data?”
“Does the server need to manage this too?”
Answering them requires understanding the service’s purpose, user flows, and scope of responsibility. Giving an AI enough context can lead to better suggestions, but developers still need to decide what to build and where its responsibilities end.
Data, in particular, is an asset that accumulates over time. If you collect it using the wrong structure from the start, rebuilding it is not as simple as rewriting code. That is why I review the DB schema myself, even when I delegate client and server coding to AI.
What is the first thing you would check in the DB schema of the service you are building?
This article summarizes my Korean YouTube video (English subtitles available) and was translated with AI. #ABotWroteThis





Top comments (0)