Rui Faria

UX/UI & Web Designer

Automating the slowest part of requirements engineering

Problem

Authoring low-level requirements in safety-critical systems was consuming disproportionate engineer time. Tighter delivery deadlines made the existing manual process unsustainable.

Approach

Design and build an AI-assisted tool using prompt engineering to generate low-level requirements from either source code or high-level requirements. Validated through contextual inquiry with practicing engineers.

30%

Increase on average
workflow improvement

4/10

Engineers adopted
the tool post-study

80%

Review
pass rate

In safety-critical software, requirements aren't documentation. They're the contractual baseline every line of code is traced against.

Low-level requirements function as verbose pseudocode. Engineers either deduced them from high-level requirements or reverse-engineered them from existing code. Both approaches were accurate, but authoring and subsequent reviews were time cumbersome. With tighter delivery deadlines arriving, the bottleneck was no longer acceptable. The tool was conceived as an internal first iteration, a proof of concept with a small group of engineers who knew the requirements process well enough to validate whether AI-generated output could meet the baseline standard.

01. Discover

Understand where time was actually going

Contextual inquiries and interviews with practicing engineers revealed two distinct time sinks: Authoring, writing requirements from scratch and Reviewing, particularly when requirement guidelines shifted mid-project and previously authored requirements needed revisiting. Authoring was the primary target. A single engineer could spend a disproportionate portion of their sprint on requirements alone, with review cycles adding further delay when changes were requested.

A beige living room corner faded background and a koi fish logo in the forefront
A beige living room corner faded background and a koi fish logo in the forefront

02. Define

Frame the right problem before building anything

The temptation was to build a writing assistant. However, engineers weren't slow writers, they were slow because every requirement needed to be individually deduced from either a high-level specification or from code behaviour. The tool's job was to generate low level requirements in bulk,a complete draft that an engineer could review, correct, and send a polished version for review. This would shifting the cognitive load from authoring to validation and speed the process when handling multiple requirements.

03. Design

Two input modes, one consistent output standard

The tool accepted two input types: source code files, which it would parse to generate requirements describing the code's behaviour; and high-level requirements documents, from which it generated the corresponding low-level breakdown. Through few-shot prompting we defined a system prompt with general guidelines and examples. Requirements had to conform to a specific structure, vocabulary, and traceability conventions that the guidelines specified. Getting the prompt architecture right took more iteration than the interface itself.

A beige living room corner faded background and a koi fish logo in the forefront
A beige living room corner faded background and a koi fish logo in the forefront

04. Validate

Measure time on task, not subjective satisfaction

Ten engineers with strong requirements knowledge were selected for the validation group. Time on task was measured before and after. How long did it take to produce a set of requirements manually versus reviewing and approving the tool's output? The 30% efficiency improvement came from that direct comparison. The validation group was intentionally knowledgeable, which means the gains likely represent a best-case scenario rather than an average across the broader team.

30%

Increase on average
workflow improvement

4/10

Engineers adopted
the tool post-study

80%

Review
pass rate

Room from improvement and next iterations.

Ten engineers with strong existing knowledge is a proof of concept, not a validated rollout. The 30% improvement is real within that group, but these were participants who could identify a subtly wrong requirement immediately. Extending the test group with engineers of varying experience levels would likely show smaller gains and surface fail points the current validation missed. The deeper open question is review confidence. In safety-critical systems, a requirement that is efficiently generated but subtly incorrect is worse than a slow, accurate one. The tool improved authoring speed, but the cognitive burden of trusting AI-generated output in a domain with zero tolerance for ambiguity wasn't fully resolved. A next iteration would need structured validation steps built into the tool itself. Mechanisms that help reviewers catch inferred behaviour before it enters the submitting the requrirements. With recent advancements in AI, a larger context window is possible, so adding the requirement guidelines as context as well as more examples, could yield a more refined output. The potential showed itself and the ambition of growing to a full-fleshed tool to handle most of the authoring was present, but not yet attainable.