Habr AI explained why Zero Shot is dangerous for deriving requirements from laws and regulations

On Habr AI, there was an analysis of why attempting to extract IT system requirements from a law with a single zero-shot prompt almost always produces false confidence. The author argues that the problem is not in the model's ability to read regulatory documents, but in the fact that between a legal norm and a system requirement lies an entire chain of analytical decisions.

Why a Law Is Not a Specification

The idea seems logical: a law already contains rules, a model can read text and structure it, so it remains only to ask to "extract requirements." But this is precisely where the substitution occurs. A regulatory document describes not an interface, process, or API, but a legal field: roles, conditions, definitions, prohibitions, consequences, and references to other norms.

Translating all of this into project requirements in one step is impossible without losing meaning. For an analyst, this is merely raw material, not a finished problem statement for the development team. The author shows that Zero Shot prematurely makes it appear as though the work is already complete.

The model collects obligations of different participants into a single list, mixes definitions, constraints, and actions, then packages everything in a convenient "system must" format. On reading, such an answer looks coherent, but it fits poorly into architecture, backlog, and verifiable specification, because it does not answer basic questions: who acts, what exactly happens, and in what part of the system should this live.

Where Losses Occur

The main complaint about the approach is not simply inaccuracy, but loss of manageability. When a team receives a ready-made list of "requirements," it becomes difficult for them to understand what exactly the model discarded along the way, what it interpreted itself, and which norms actually relate not to the product, but to external participants in the process. This is how the boundary between a mandatory requirement, a working hypothesis, and simply a restatement of a norm is lost. This creates a dangerous illusion of completeness.

• Exceptions and conditions of norm application may disappear from the response
• Constraints are easily transformed into features
• Definitions are masked as full-fledged requirements
• The subject of action is lost: who initiates, verifies, and records the result
• Traceability from a specific norm to a project conclusion is lost

Because of this, a polished answer turns out to be unverifiable. If someone on the team asks where a specific point came from, a reference to the law article will be insufficient. Traceability is needed: what fragment of the norm produced what conclusion and why. Without it, any requirement is vulnerable to discussion with lawyers, analysts, and developers, especially if it comes to complex scenarios, audits, or regulatory inspections. And the stricter the regulatory environment, the more costly such lack of transparency becomes.

Where Zero Shot Is Useful

At the same time, the author does not propose abandoning LLMs in work with legislation. On the contrary, Zero Shot can be useful as an initial reconnaissance: quickly get into the topic, obtain a draft map of entities, throw out hypotheses about scenarios, and understand which parts of the document require manual analysis first. The problem begins the moment this draft starts being treated as a final result of the analysis. As a tool for primary text navigation, this saves time.

"Zero Shot can be used to start work with a law.

But it should not be used to consider that work complete."

A working alternative looks less striking, but is more reliable: first determine for which role and which object of automation the requirements are being gathered, then break down the text by types of material, restore subjectivity, check completeness, and only after that transform the conclusions into project artifacts. AI in such a scheme remains a useful assistant, but no longer substitutes for the analytical logic itself. It is in such a mode that a model accelerates work without substituting expertise.

What This Means

For teams that want to accelerate legal analysis with the help of LLMs, the conclusion is simple: one polished prompt does not replace a systematic analysis of regulatory documents. Zero Shot is suitable for the first pass and drafts, but not for defendable requirements upon which product decisions, architecture, and legal compliance depend. Otherwise, speed at the input turns into errors already at the stage of actual design and coordination, and not only in theory.