Session 05 / AI-native engineering
Spec-Driven
Development
in the AI age.
When implementation is abundant, a clear specification becomes the shared contract.
Session 05 / AI-native engineering
When implementation is abundant, a clear specification becomes the shared contract.
No one codes alone.
The shift
Implementation moved downstream.
Choose the mode
Shift gears as the cost of being wrong rises.
One app / two processes
Decide inside the loop
Decide before the build
Same app. Different location for engineering decisions.
What a spec is
Example / Community event registration
Clear enough to build. Specific enough to verify.
The artifact chain
GitHub Spec Kit, simplified.
Live demo
The spec removes decisions from the coding loop.
Make it testable
100 people are confirmed
A new person registers
They join the waitlist
If “done” is observable, the agent can verify it.
How teams use it
Humans decide. Agents accelerate. Evidence closes the loop.
Tool landscape
| Tool | Core bet | Best fit |
|---|---|---|
| Kiro | Specs live inside the development environment | Guided feature work |
| Spec Kit | An open artifact workflow can work across agents | Teams that want portability |
| Tessl | The specification can become the source | Higher automation experiments |
The workflow matters more than the logo.
Maturity
Write the spec before implementation.
Keep specification and code aligned as both evolve.
Humans change the spec; the system maintains the code.
Most teams are still learning Level 1 and Level 2.
What the evidence says
A machine-readable security constitution constrained AI-generated banking services.
One case study / preprintRepository discovery hooks reduced context blindness and improved benchmark results.
Early evaluation / preprintUse the findings as signals, not guarantees.
What SDD does not fix
A detailed specification can precisely describe the wrong product.
The document drifts away from the code, users, or constraints.
Generated code looks complete without proving the behavior.
Ground it. Version it. Test it. Review it.
The decision rule
Tiny, reversible experiment with no important data.
Users, data, several decisions, or team handoffs.
Payments, privacy, security, or irreversible actions.
More consequence means more explicit agreement.
Before the next agent run
No one codes alone. oslovibecoding.tech