TNO070: Spec Driven Design (SDD) for NetOps and NetEng

The episode features an interview with Lasse Haugen, a network operations engineer at Norway's public road administration, recorded at AutoCon 5 in Munich. Haugen began his career in network automation by necessity—setting up computer parties at age 14 required rapid configuration deployment. This led to work on The Gathering, Norway's largest computer party with 5,000+ participants, where he developed zero-touch provisioning solutions. He now works in a specialized role focused entirely on automation, freed from day-to-day outage response through team collaboration.

The core topic is Spec-Driven Development (SDD), a methodology that uses AI to create disciplined, guardrailed automation projects. Rather than using AI for direct code generation (vibe coding), SDD involves defining intent, constraints, and success criteria in structured specifications. Haugen describes his journey from initial skepticism about AI tools to trusting them through iterative specification development. His primary example is Taylite, a syslog collector project that underwent four iterations of specification refinement, ultimately reducing production logs by 98% to surface only operationally relevant messages.

Haugen's workflow demonstrates how SDD differs from traditional vibe coding: code is treated as cheap and disposable, with the specification as the valuable artifact. He moved beyond manual spec creation to having AI generate specifications based on his intent, then used Claude's "Grill Me" skill to identify overlooked requirements and edge cases. For a Google Cloud infrastructure project, Claude conducted six rounds of interview questions about his two-datacenter topology, automatically identifying the right SR Linux and Junos images, validated designs, and Terraform configurations.

Key practices include treating specifications as product requirements documents that non-technical stakeholders can review, using AI as a collaborative thought partner rather than a code generator, and building trust through repeated iteration and testing. Haugen demonstrates productivity gains through unconventional methods—walking in forests while discussing project architecture with Claude, dictating notes while driving, and using a "brain dump agent" to automatically organize and categorize conference learnings into structured project documentation.

For his next major project, a GNMI-based in-house monitoring solution, he's generating a 10,000-line specification entirely through dialogue with Claude before any code development begins. He emphasizes that AI excels at identifying database design patterns and operational considerations that non-DBAs might overlook. Haugen also discusses trust-building: starting with small experiments, accepting that specifications will evolve, maintaining conversation history for context, and leveraging AI's memory of previous projects to maintain consistent implementation patterns.

The conversation touches on cost considerations, suggesting that as token costs rise, local models will become more attractive. Haugen contrasts SDD with traditional SaaS tools, noting that in-house solutions developed with AI allow for immediate feature implementation without vendor roadmap delays. His worst outage was taking down internet connectivity at The Gathering due to local ISP configuration issues—a cautionary tale about infrastructure dependencies at scale.