How a Single System Prompt Turns an LLM into a Reliable Tool: Tests on Qwen and DeepSeek

Large language models lie beautifully. Not out of malice: they are trained to continue text, not to tell the truth. Where the model lacks the necessary data, it generates something plausible and delivers it with an expert's confidence. For applied tasks such as corporate assistants, analytics tools and decision-support systems, this behavior is unacceptable. A confidently delivered error is worse than an error with a caveat.

The Exoskeleton Metaphor

The author of an article on Habr proposed a simple but workable metaphor: an LLM needs an exoskeleton. Not additional training, not an RLHF round, not expensive fine-tuning, but a single system instruction that sets strict behavioral rules for situations of uncertainty. The tests ran on two of the most popular open models with strong Russian-language support: Qwen (Alibaba's series) and DeepSeek. Both are widely used in Russian products precisely because of their availability and quality.

The Core of the Approach: Behavioral Rules

The point of the "exoskeleton" is to forbid the model from being overconfident where it is not sure. The system instruction prescribes several key rules. First: acknowledge uncertainty explicitly.

Do not stay silent about it; say outright "I don't know" or "I don't have enough data." Second: clarify the request if it is ambiguous, instead of picking one interpretation and answering that. Third: clearly separate facts the model is sure of from those it merely assumes.

Fourth: decline to answer in areas where the risk of error is high and the information cannot be verified from inside the model. In practice this sounds trivial. In practice, it works.

After the instruction was added, Qwen and DeepSeek began to acknowledge the limits of their knowledge much more often: in test scenarios with deliberately insufficient or contradictory context, the models stopped "filling in the blanks" and started requesting clarification or explicitly flagging uncertainty. The rate of confident hallucinations in these scenarios dropped noticeably.

Why This Isn't Obvious

Why isn't this obvious? Because by default an LLM is trained to give a complete, confident answer, and that is what earned it high scores in RLHF. A human rater instinctively prefers a detailed, confident text to a short "I don't know."

The model absorbed that preference. As a result, it has built-in behavior that is the exact opposite of what real production needs, where the cost of an error is measured in reputation or money. A system instruction is a way to overwrite this behavior without changing the model's weights.

In essence, we impose epistemic humility on the model from the outside. Hence the exoskeleton metaphor: the model itself does not change inside, but a rigid structure of behavior forms around it and steers its reactions in the right direction.

The Instruction Must Be Specific

An important nuance: the instruction must be specific, not declarative. "Be accurate and honest" does not work, because the model already considers itself accurate and honest. What works are concrete situations: if the request lacks sufficient context, ask a clarifying question; if you are not sure of a fact, say so explicitly and explain why; if the question goes beyond your data, say so directly. Each rule describes a specific trigger and a specific action in response.

Does It Reduce Usefulness?

Developers often fear that restrictions will reduce a model's usefulness. The tests did not show this. In scenarios with sufficient context, the models performed as well as without the instruction.

The restriction kicked in only where data was genuinely lacking, which is exactly where the model used to hallucinate. For teams building internal tools on LLMs, such as corporate knowledge bases, analytics assistants and document management systems, this result is practically applicable right now. There is no need to wait for the next model version, allocate a fine-tuning budget or change the architecture.

It is enough to write the system prompt properly, and the model starts behaving the way the business needs, not the way it was trained to please random raters.

In brief:

• A single system instruction (the "exoskeleton") sets rules for LLM behavior under uncertainty, without additional training or fine-tuning.
• Tests on Qwen and DeepSeek: the models admitted the limits of their knowledge more often and produced fewer confident hallucinations.
• The instruction must be specific (trigger → action); in scenarios with sufficient context, usefulness did not decrease.