Bad vibe coder or good AI engineer? Four decisions that tell them apart
At Seznam, developers decide how much text to give a model. At Košík, they ask whether a colleague will still be able to follow the code a year from now. Four skills from Andrew Ng’s map come into focus in Czech software teams.

Imagine two developers receive the same brief: “Build us a customer support assistant.” Both have access to the same coding agent. The first one is showing a working chat by lunchtime. The second spent the morning with a support colleague. They found out that it takes her half a minute to write a response to a customer, but eight minutes to look up what someone else promised last week. So, what do you assign her to on Monday?
That is how I read the skills map that Andrew Ng presented on 14 August 2026. His team at DeepLearning.AI analysed over ten thousand job ads, conducted interviews with experts and hiring managers, and supplemented this with surveys. The result is four areas that make sense to focus on when developing with AI. I have selected examples from domestic job listings. They clearly highlight what a developer can already learn today, even if their job title does not contain the word “AI”. Andrew Ng on X · DeepLearning.AI Skills Map
The same agent, four different choices
| Skill | Bad vibe coder | Good AI engineer |
|---|---|---|
| Measuring quality | Vibe coderDemonstrates one successful chat. | AI engineerTests the assistant on closed cases to see how often it finds the right promise and how much time it saves. |
| Understanding the system | Vibe coderClicks “Refund”. It works. Done. | AI engineerInterrupts the connection and clicks again. Verifies that the money was sent only once. |
| Directing agents | Vibe coderTypes “fix the refund” and accepts the first adjustment. | AI engineerDescribes what should happen if the confirmation goes missing, then tests the agent’s change against that example. |
| Defining the task | Vibe coderHears “we want a chat” and starts building it. | AI engineerSits down with the support team. If finding a past promise takes the most time, first tries a link to the original message. |
Andrew Ng’s skills map in Czech teams
Andrew Ng · When developing an AI application, continuously measure how well it solves the task for which it was created.
Czech practice · For an assistant, the developer must monitor search accuracy, response speed, and computational resource consumption. Seznam
Andrew Ng · Knowledge of architecture, data, and operations allows designing a system that stands up even during an outage.
Czech practice · The developer is responsible for the service from design to operation. They must monitor security, performance, and whether a colleague can understand the code in a year. Košík
Andrew Ng · Decide what an agent can do independently and how you will verify its work before taking it over.
Czech practice · According to the job ad, an agent writes most of the code. The company expects the developer to provide the task, plan, and thorough review of proposed changes. DX Heroes, Prague
Andrew Ng · Part of development is deciding what should be created, who it will serve, and how you will recognize the benefit.
Czech practice · A business professional should map the workflow, eliminate unnecessary steps, and only then deploy agents. Groupon, Prague
1. Know how to build an AI application and measure who it helped
Take a colleague from support. The assistant now reviews the order history and finds the sentence: “We will refund your shipping.” It looks great. But the customer had two packages sent. Which one did the promise refer to? This is a much more interesting test than asking whether a chatbot can answer fluently in Czech.
At Seznam, the developer of their AI assistant is tasked with selecting the context for the answer. If the model receives too little text, it may miss important context. If you send it everything, token consumption and user wait times increase. A person in this role must be able to measure search accuracy, response time, and resource usage. How Seznam describes the role
How to start measuring? Pick twenty closed cases with someone from support. For each, pre-mark the message containing the decisive promise. Then compare two versions of the assistant. How often did they find the correct message? How long did it take to resolve the case? Those twenty examples will give you the first clue as to whether it is worth changing the search, the model, or the interface itself.
2. Understand what happens after the click
The model proposed refunding the shipping. The operator clicks “Approve,” nothing happens for a moment, and they click again. How many times does the money go out? The answer depends on how the application, database, and payment service communicate. Whoever understands that path can set a precise condition for the agent and verify it even during a connection failure.
At Košík, they plan to use AI for design, code writing, testing, and deployment. Developers, however, retain responsibility for security, performance, and ensuring that another colleague can understand their work a year later. A pretty good topic for the next code review. Could you explain to a colleague, using one real order, where they can see whether the payment was sent and whether it is safe to repeat it? What Košík requires
3. Give a coding agent work you can verify and take over
The Groupon team in Prague describes the division of labour simply. The developer prepares the specification and tests, can split the work among several coding agents, and then verifies and deploys the result. In the application, they want to see two or three things the candidate has built. For each, they must state which part they did, which part the team did, and which part AI did. How Groupon hires
You can make a big step just by changing one work habit. Before launching the agent, write an example against which you will accept its solution. For a shipping refund, for instance: “The first request went through, but the confirmation was lost. After the second click, the customer sees one refund and there is a single outgoing payment on the account.” You can leave the implementation to the agent. You have decided what promise you are making to the customer. And you have a test that will remain valid after the next model swap.
4. Know how to change the specification before it becomes software
Ng calls the fourth area shaping the build. This is where the work happens that determines what will actually be created. In our example, I would show my colleague in customer support two options. One would draft a polite response. The other would open the correct message, highlight the original promise, and let her answer in her own words. I would have her handle several requests both ways. She might choose the simpler application. Or we might discover that the most time-saving solution is a fairly ordinary link from the order to the correspondence.
I like vibe coding for moments like this. Two variants can be built quickly and put into people's hands while it is still cheap to discard one. A developer distinguishes themselves by what they learn in the process. How many minutes is it worth saving? Where does the customer need an immediate reaction and where would they rather wait for verification? An agent can calculate the cost of both variants. To make a choice, it also needs to know the priorities of the people who will work with the result.
Two sentences before the next prompt
You can try the same thing outside IT. Ask a chatbot for a budget for a new corporate website and you will easily get a neat table with the price of design and implementation. If you have to decide between two suppliers for three years based on it, you also need hosting, maintenance, and the time of your own staff included.
Try writing two sentences before a task for AI today. The first states what decision you will use the output for. The second defines what must not be sacrificed in that process. For a budget, I would write: “I want to choose a supplier based on total costs over three years. Include operations in the comparison and explicitly mark what we need to find out for items the supplier did not price.” Now it makes sense to open the chat.
For me, the sign of a good AI engineer is that satisfied customers come back even when they have other options. A good engineer can deliver faster, at lower cost and more reliably than the competition. What matters most is that the solution works well, is secure and people enjoy using it.
Which compromise in your work can AI not estimate for you on its own?
Related articles
May 2026
Kasandra Superforecaster: an AI salon where the future is not guessed by feeling, but forecast carefully
ReadOctober 2026
Shadows of Prague: six AI minds, two rival teams, and the journey from blocks to game characters
ReadOctober 2026