Skip to content

Spec-driven AI Development

Specification as source of truth

Problem

AI can generate code extremely fast, but without direction it creates uncontrolled output: code that doesn’t align with requirements, isn’t reviewable, and that no one takes responsibility for. Speed without direction is not progress.

Approach

The specification steers the process. KIRO steering documents define intent, requirements, design and tasks before AI begins implementing. Every step is reviewable, traceable and verifiable. AI proposes and accelerates; humans specify, review and decide.

Process

  • Intent — What do we want to achieve and why?
  • Requirements — Define acceptance criteria and constraints
  • Design — Document architecture and technical decisions
  • Tasks — Create implementation plan with traceability to requirements
  • Implement — AI generates, humans review and validate
  • Validate — Run tests, evidence and acceptance checks
  • Evolve — Iterate based on feedback and new insights

Controls & evidence

  • Traceability — Every implementation traces back to requirements, design decisions and tasks. Nothing is built without a documented reason.
  • Reviewable increments — Every AI-generated step is small enough for human review. No black-box bulk generation.
  • Security checks — Security requirements are explicitly specified and validated, not implicitly assumed.
  • Validation evidence — Every requirement has measurable proof that acceptance criteria are met.

Outcome

Software that demonstrably meets requirements, is traceable from intent to implementation, and where every AI contribution is human-validated. Faster delivery without loss of control or quality.