LMU Media Informatics Group
Research Project at the Media Informatics Group at LMU funded by bidt

Human-Centered
Specification-Driven
Software Engineering
with Generative AI

What if anyone—scientist, engineer, artist—could build reliable software simply by describing what they need? AI-SPEC2SW researches how generative AI can guide non-experts through creating precise, human-readable specifications that translate automatically into working software.

Project ID
AI-SPEC2SW
Funding
bidt (Bayerisches Forschungsinstitut für Digitale Transformation)
Duration
01/2025 – 12/2028

The Challenge

Who gets to shape digital reality?

Software has become fundamental infrastructure—yet only a small fraction of society has the skills to create it. Everyone else depends on tools built by others, with no ability to adapt them to their needs. AI-SPEC2SW is built on the premise that this is neither inevitable nor acceptable.

The critical insight: rather than generating code directly from vague prompts—an approach that produces opaque, brittle, and hard-to-verify results—we use generative AI to guide users through building a structured specification. This specification is human-readable, verifiable, and complete enough to generate reliable software and tests automatically.

Our Approach

Why current AI code generation is not enough

Current Limitations

Problems with direct AI code generation

  • Users must understand how to deploy, run, and integrate generated code
  • Generated source code is opaque—users cannot verify it does what they expect
  • Vague or incomplete prompts lead to hallucinated functionality
  • Projects beyond simple snippets require software architecture skills most non-experts lack
  • Debugging and testing requires expertise that end-users rarely possess

The Specification Approach

How AI-SPEC2SW addresses these problems

  • AI guides users to produce a structured, human-readable specification—not raw code
  • The specification is transparent: users can read, discuss, and validate it themselves
  • Completeness and consistency are checked interactively before generation begins
  • From a complete specification, both executable software and tests are generated automatically
  • Changing requirements are handled by refining the spec, not hunting through code

Research Questions

What the project investigates

Three interconnected research questions guide our empirical and constructive research, grounded in human-centered design methodology.

RQ 1

Human-centered design in software specification

How do human-centered design principles affect the inclusivity, clarity, and completeness of software specifications produced by non-experts? How does end-user involvement in an iterative specification process affect the quality of the resulting software?

RQ 2

Generative AI for specification creation and validation

Which AI models and interaction paradigms most effectively support users in building accurate, consistent specifications? How must existing end-user development approaches be adapted—or replaced—when combined with generative AI?

RQ 3

Competencies and digital literacy

What level of initial skill is actually required to use specification-driven tools effectively? How does sustained use of these tools develop users' own abilities in software specification, and does it contribute to broader programming literacy?

Publications & Presentations

Research outputs

CHI 2026 — ACM Conference on Human Factors in Computing Systems

From Throw-Away to Takeaway: How GenAI and Vibe Coding Accelerate Prototyping Across Technical Skill Levels

Charlotte Kobiella, Daniela Breidenstein, Albrecht Schmidt · Article 688, 1–23 · DOI: 10.1145/3772318.3790757

EISEAIT Workshop @ EICS 2025 — Trier, Germany

AI-Assisted Specification-Driven Software Creation

Thomas Weber, Sven Mayer, Albrecht Schmidt

AI for EUD Workshop @ IS-EUD 2025 — Munich, Germany

Empowering the End-User — Software Development with LLMs

Thomas Weber, Passant Elagroudy, Philippe Palanque, Sven Mayer

bidt-Konferenz 2025 — Invited Talk · October 16, 2025

AI-SPEC2SW — Menschzentrierte, spezifikationsgetriebene Softwareentwicklung mit generativer KI

Albrecht Schmidt

Team

People

Albrecht Schmidt

Prof. Dr. Albrecht Schmidt

Principal Investigator · Chair for Human-Centered Ubiquitous Media · LMU München

Thomas Weber

Dr. Thomas Weber

Postdoctoral Researcher · Chair for Human-Centered Ubiquitous Media · LMU München

IW

Isabel Wanderwitz

Doctoral Researcher · Chair for Human-Centered Ubiquitous Media · LMU München