Spec-driven development (SDD)
Spec-driven development (SDD) is an emerging family of AI-assisted software workflows in which a structured specification guides planning, task decomposition, implementation, and verification. Tools commonly turn stated behavior, constraints, and acceptance conditions into plans, tasks, and code. Interpretations differ: some use a specification mainly to start implementation, while others seek to maintain it as the primary source of intent. The term does not guarantee that every workflow keeps the specification and code synchronized.
Origin and context
Kiro introduced its spec-driven development workflow in July 2025. GitHub's Spec Kit history dates its initial public work to 21 August 2025 and documents a structured command flow for specifying, planning, and implementing software. Thoughtworks Technology Radar later placed SDD in Assess, described competing interpretations, and noted tools including Kiro and Spec Kit. A 2026 preprint analyzed SDD as a contemporary AI-coding practice while relating it to older specification-first traditions.
Why it matters
A shared specification can make product intent visible before an agent produces a large patch. It gives reviewers an artifact against which to question assumptions, align acceptance criteria, and trace tasks. For multi-step agent work, the spec can also reduce reliance on a transient chat history. The benefit depends on the quality and maintenance of the artifact: a precise-looking document can preserve a mistaken requirement just as efficiently as a correct one.
Example
Before asking an agent to add notification preferences, a team records supported channels, default states, save behavior and acceptance examples in a versioned specification. The tool derives a technical plan and implementation tasks from that artifact. Review checks both the changed code and the intended behavior. If the feature's requirements change, the team updates the specification as well. This illustrates a maintained spec-to-implementation workflow; a one-line request to add settings lacks that explicit artifact and traceable decomposition.
How it differs
Evaluation-driven development (EDD)
SDD organizes development around an explicit specification of intended behavior and constraints. Evaluation-driven development organizes iteration around cases, criteria, and scores that test observed system behavior. A team can derive eval cases from a spec and use both practices, but a specification is not itself evidence that the implementation meets it.
AGENTS.md
An AGENTS.md file gives repository-wide or directory-scoped operating instructions to coding agents, such as commands and conventions. An SDD specification describes the intended behavior and constraints of a particular feature or system. Repository guidance can tell an agent how to execute an SDD workflow, but it is not the feature specification.
Maturity and evidence
Maturity is rated 3. Several independent tools now implement recognizable spec-to-plan-to-task workflows, and independent analysis gives the label a coherent practical scope. The lifecycle remains emerging because definitions range from lightweight spec-first development to treating specifications as the primary executable artifact. Evidence that these workflows reliably improve quality or productivity across teams is not yet strong enough for a higher rating.
Limits and open questions
Up-front specification can be disproportionate for small or exploratory changes, and generated documents can create review burden without adding shared understanding. Ambiguous or incorrect specs can steer an agent consistently in the wrong direction, while code and spec can drift after delivery. Tool-specific command sequences are not a universal method. Teams should match rigor to risk, assign ownership, record open questions, validate critical assumptions, and retain conventional code review, testing, and production feedback.
Related terms
References
- Introducing KiroKiro · 2025-07-14 · class A
- HistoryGitHub · 2026-08-21 · class A
- Technology Radar Volume 33Thoughtworks · 2025-11 · class B
- Spec-Driven Development: From Code to Contract in the Age of AI Coding AssistantsarXiv · 2026-01-30 · class B
- Iterating Towards LLM Reliability with Evaluation Driven DevelopmentLangChain · 2024-03-11 · class A
- AGENTS.mdAGENTS.md Project · 2025-12-10 · class A
Last updated: 2026-09-05
This term is also covered in the Skills Atlas as ai requirements engineering skill.
This term is also covered in the Skills Atlas as ai assisted development skill.