A Hands-on Report on Agentic Coding with SDD (Spec-kit)

While building an AI platform, the productivity of agentic coding was strongly felt, but as the project grew more complex, indiscriminate code generation and hallucination emerged as problems. So the Kakaopay AI Platform team introduced SDD (Spec-Driven Development) and Spec-kit, shifting the developer's role away from writing code and toward spec definition, verification, and test design.

In Spec-kit, you first set the team's top-level principles with constitution, then flesh out feature requirements with specify. If needed, you further clarify requirements with clarify, then create a development plan in plan and break it down into executable units of work with tasks. Finally, analyze checks for contradictions among the spec, plan, and tasks, after which code is generated in the implement stage.

For setup, install the CLI with uv tool install specify-cli --from git+https://github.com/github/spec-kit.git, then get started with specify init <PROJECT_NAME> or specify init --here. After initialization, if you request that it proceed in Korean, artifacts such as constitution.md, spec.md, plan.md, and tasks.md are created in sequence, and external context such as Figma MCP can also be connected.

The most important point in practice is that rather than directly reworking a result you don't like, you go back to specify and fix the spec there. Doing so reduces inconsistencies among the spec, plan, and code, and at the team level, it also makes it easier to track what other developers did through documentation.

However, the many steps make it complex at first, it can actually be a burden for simple fixes, and the amount of generated documents and token usage also increases. Even so, because elements that must not waver—such as security, testing, and consistency—can be managed structurally, the conclusion is that SDD is valid as the next step for agentic coding.