Spec-Driven Dev Is Back. But Not How You Think

Spec-Driven Dev Is Back. But Not How You Think

by Daniel Terhorst-North & Gojko Adzic

Applied Sciences & Engineering Computing & Cybersecurity UMComputer programmingUMFAgile programming
🎙 Daniel Terhorst-North & Gojko Adzic 👥 1.1M 📅 April 27, 2026 ⏱ 39 min 👁 5K 📄 expert opinion 🧭 2026-08-02
Available in: English (current) Français

Keywords

spec-drivenAI agentsiterative developmentlintingguardrails

Summary

In this GOTO Unscripted interview, Daniel Terhorst-North and Gojko Adzic discuss the resurgence of spec-driven development in the context of AI-assisted coding. They argue that one-shot ‘spec-to-code’ approaches are doomed to fail, drawing parallels to CASE tools and model-driven architecture of the past. Instead, they advocate for iterative, feedback-driven loops where AI handles deterministic coding tasks while humans maintain domain knowledge and architectural judgment. Gojko shares a practical solution for enforcing domain-specific rules: encoding them as automated linting checks (e.g., custom ESLint rules) that run in CI, applicable to both humans and AI agents. They emphasize that AI lacks semantic understanding and must be constrained by automated guardrails rather than markdown files. The conversation also touches on the commercial motivations behind these tools and the importance of treating AI as a ‘stochastic parrot’ rather than a sentient being. They conclude that while AI can accelerate development, it does not replace the need for human oversight and quality assurance.

158 words

Critical Evaluation

The interview provides a valuable and nuanced perspective on the current state of spec-driven development with AI. Both speakers are highly credible, with Daniel Terhorst-North being the originator of BDD and Gojko Adzic a well-known consultant and author. Their arguments are well-structured and grounded in historical precedents, such as CASE tools and MDA, which strengthens the analysis. The discussion is practical, offering concrete advice on how to effectively use AI in development, particularly the idea of encoding domain rules as automated linting checks. This is a novel and actionable insight that goes beyond common advice. However, the conversation is largely anecdotal, relying on personal experiences rather than empirical data. The speakers do not cite specific studies or metrics to support their claims, which limits the scientific rigor. Additionally, while they critique the limitations of AI, they do not delve deeply into potential solutions for improving AI’s semantic understanding. The adéquation between title and content is strong, as the title accurately reflects the discussion. Overall, the interview is insightful and thought-provoking, but its reliance on anecdotal evidence and lack of empirical support prevent it from being a definitive scientific resource. The public comments, if any, were not provided, so no analysis of audience reception is included.

205 words

Title / Content Match

The title accurately reflects the content, which discusses the resurgence of spec-driven development in the context of AI, arguing against one-shot approaches while advocating for iterative, feedback-driven methods.

Quality & Reliability

8/10

The speakers are recognized experts in software development, with deep practical experience. The discussion is nuanced, critical, and grounded in historical precedents, though it lacks empirical data and relies on anecdotal evidence.

Chapters

Cited Sources

Concurring Sources

  • Martin Fowler's Article on Code as Design — Martin Fowler's argument that programming is about human understanding, aligning with the speakers' view.

Dissenting Sources

  • Spec-Kit by Microsoft

External References

Contribution & Novelties

The interview offers a fresh perspective on spec-driven development in the AI era, emphasizing the importance of iterative feedback loops and automated guardrails. The key novel insight is Gojko’s approach of encoding domain-specific rules as custom linting checks, which provides deterministic enforcement for both human and AI developers. This moves beyond the common reliance on markdown files and offers a practical, scalable solution.

Pour aller plus loin :

  • Behavior-Driven Development (BDD) — BDD is a methodology that encourages collaboration and specification by example, relevant to the iterative spec-driven approach discussed.
  • ESLint — The tool used to implement custom linting rules, enabling automated enforcement of coding standards.
  • Stochastic Parrots (Bender et al., 2021) — The paper coining the term ‘stochastic parrot’, which the speakers reference to highlight AI’s lack of understanding.

130 words

Radar Profile

The radar profile shows high scores in information quantity, quality, and reliability, with a slightly lower technical depth. This indicates a well-rounded discussion that is both informative and credible, though it may not delve into highly technical details.

Reliability 8/10