← Articles

Why more instructions made the conversation with AI worse

5 min read (updated: August 24, 2026)

The first version worked — until it started talking to people

The assistant on my site is meant to help a business owner describe a problem before getting in touch. It should not pretend to be a human, promise a price or stage an interrogation made up of technical questions.

The first tests quickly revealed small errors. Once the model spoke about me in the first person. Another time it asked three questions at once. It could also move to the summary too early, or keep asking about technology although the client was simply describing a mess in their orders.

For every case I added another rule. Soon the instruction had a dozen or so prohibitions, exceptions and conditions.

Each one sounded reasonable on its own. Together they made the conversation worse.

Symptoms of an overloaded instruction

The model started to watch the rules more than to listen to the person. The answers were correct, but stiff. When a client strayed from the path I had foreseen, the assistant tried to push them back into its plan.

A conflict of goals appeared as well. It was supposed to be brief, but thorough. It was supposed not to guess, yet at the same time propose the next step. It was supposed to collect contact details, but not to press. The more such sentences I added without setting priorities, the less predictable the result was.

This is a common mistake when working with language models: a bad answer looks like one missing rule. Meanwhile the cause may be too many rules, or a wrong division of responsibility.

Changing the model did not fix the design of the conversation

I tried a different model. I hoped it would cope better with a long instruction and keep more freedom. Some of the answers did sound more natural, but new problems arose: the model added facts the client had not given, or ended the conversation at a moment when it should still have waited for an answer.

That was an important signal. The problem did not lie only in the choice of tool. The design itself mixed together too many tasks.

I went back to the previous version and simplified the whole thing.

What works better

Hard limits belong in the code

The message length limit, the number of conversations from one address, the session’s period of validity and the saving of data should not depend on the model’s good will. The program checks them itself.

The end of the conversation works similarly. The model can suggest that the material is complete, but it is the application that removes the technical marker, saves the state and shows the client the way forward.

A conversation needs a few clear rules

After the simplification, the most important rules remained:

  • introduce yourself as Artur’s assistant, not as Artur,
  • ask one question at a time,
  • ask about how the company works before moving on to technology,
  • do not give a price or a deadline without a human’s analysis,
  • do not ask twice for information that has already been given,
  • finish when the material is enough to choose the next step.

It is still a specific instruction. The difference is that it describes the goal of the conversation instead of trying to foresee every possible sentence from the client.

The summary is a separate task

The model running the conversation does not have to create a report after every answer at the same time. Only at the end does a second step arrange the data into fixed fields: the problem, the current way of working, the people, the systems, the risks and the expected result.

Separating the conversation from the organizing made both tasks simpler.

Better to describe uncertainty than to forbid everything

Instead of forbidding any suggestion at all, the assistant can say: “This is one of the possible routes; Artur will assess it once he knows the data and the constraints.” The client gets a useful direction, but does not take a loose hypothesis for an offer.

How I test the conversation

The question “does the answer sound good” is not enough. I check whole scenarios:

  • an owner describes the problem without any names of technologies,
  • a technical person gives a lot of detail from the first sentence,
  • the client changes the subject halfway through,
  • an important piece of information appears only towards the end,
  • someone gives an e-mail address early, or does not give it at all,
  • the inquiry does not concern the services at all.

I look at whether the assistant avoids repeating questions and adding facts, and whether it ends with a summary on the basis of which a decision can really be made.

The most important takeaway

Designing a conversation with AI is not about finding one magic command. It is a combination of a clear goal, limits written into the program, a short instruction and trials on real ways of speaking.

Sometimes the best fix is to remove five earlier fixes.

Check the current version of the assistant.

Have a similar problem?

Describe it in your own words — you do not need to know the names of technologies or prepare a specification.

Describe the problem