From Specs to Code with AI Agents — Simon Martinelli's practical guide to making specifications the single source of truth for code, tests, and validation, built on the AI Unified Process.
Code-centric development leads to maintenance problems, hinders modernization, and causes business misalignment.
Traditional development is code-centric. Requirements get outdated, documentation drifts, and when bugs appear, we dig through code to understand what the system was supposed to do.
AI coding tools make this worse by generating code faster without fixing the underlying process problems. Even spec-driven tools mostly use the spec to describe the code a developer is about to write — the code stays the real artifact, and the spec is just a means to produce it.
AI Unified Process flips this around. Requirements stay at the center, and everything else gets generated from them using AI as the consistency engine.
Iterative Improvement: Through short iterations, specifications, code, and tests improve together. Documentation enables sustainable development and modernization.
Test-Driven Consistency: Tests ensure the system behaves the same regardless of code generation changes, enabling safe refactoring and evolution.
In most spec-driven tools, developers write specs to describe the code. AI Unified Process specs describe the behavior of the system — the code is just the current implementation. Spec-driven development where the spec is a requirement, not a code plan.
Developer-centric by design. Specs are written by developers to describe the code: an elaborate prompt for one feature's implementation — and effectively discarded once the code ships.
The spec serves the code: the code remains the source of truth, the spec decays the moment it ships, and requirements knowledge never leaves the development team.
Requirements-centric from the start. Use cases specify the behavior of the system — what it must do, not how the code does it. They are first-class, living artifacts maintained by Requirements Engineers — or by developers deliberately doing requirements engineering.
The code serves the spec: behavior is the source of truth; code, tests, and documentation are generated — and regenerated — around it. Use cases outlive any implementation.
For us at WBS GRUPPE, the AI Unified Process is the ideal method for analyzing and evolving existing products in a structured way. With requirements and use cases consistently at the center, we stay in full control even in an established codebase. Particularly remarkable: thanks to the precise specifications, the subsequent implementation was surprisingly fast. A highly efficient framework for agile software development in the AI era.
Using AIUP, we delivered our internal B2B sales app in a fraction of the time I expected, including its interfaces to the aging in-house system that manages our product licenses. We had our old, manually maintained Word offers and sales statements analyzed and specified in markdown, and it was astounding to see the generated PDFs come out perfect, with every calculation done automatically. The specifications carried the whole thing. Now we can move functionality out of that hard-to-maintain legacy system, step by step.
I used AIUP for both a green-field and an existing brown-field project. Both worked, and both approaches were as efficient as they were fun to use. Moreover, for the first time in agentic engineering, it felt like control was given back to me, the developer — a feeling I had missed since leaving the AI-prompting phase and venturing into 'vibe-coding'. Because the aiup-vaadin-jooq tech stack was not feasible for my existing brown-field project, I contribute to AIUP by providing the aiup-angular-jpa plugin.
Building software with AIUP and Claude Code is genuinely fun. Two Rust projects came out of it: AudioSnip, a cross-platform desktop app built with Tauri 2 to extract audio from video files, and Konzertmeister CLI, a tool for the Konzertmeister API. Thanks to Claude Code, both were packaged as a Homebrew Tap, so I can install them directly with brew install on my Mac. Writing specs, implementing, and testing together with Claude Code — cool stuff. And afterwards you actually understand what the code does.
With AIUP and spec-driven development, I shipped a complete product — deckweaver — in three calendar days, with maybe four to five hours of actual work. From a two-sentence README, Claude Code generated requirements and use cases that matched exactly what I had in mind. It then handled the tedious parts — OAuth, the Google Slides API, the Thymeleaf frontend — without a hitch. Genuinely impressed.
Pick a thread, the methodology, the enterprise story, the videos, the tools, or the articles.
Four agile phases, two workflows (Greenfield and Brownfield), six core principles, and the iterative approach that replaces the determinism fallacy.
Governance and traceability, brownfield modernization, parallel team scaling, risk-managed AI evolution, and knowledge that outlives teams.
Conference talks, walkthroughs, and methodology overviews showing the AI Unified Process and spec-driven development in practice.
Open-source Claude Code plugins for the AI Unified Process workflow, plus an IntelliJ plugin that links use case specs to their tests.
Curated writing on spec-driven development, requirements engineering, AI-ready architectures, and the AI Unified Process methodology.
The AI Unified Process combines the best of proven methodologies with modern AI tooling.
Schedule a 30-minute call to discuss how the AI Unified Process fits your team and product.
Schedule a callUpdates on the AI Unified Process — methodology, tools, and case studies. No spam.
