Building the software factory from the spec up | LinearB's Ariel Illouz & Ofer Affias

The podcast discusses the implementation and impact of spec-driven development powered by AI at Linear B, highlighting a shift from informal "vibe coding" to structured, AI-assisted workflows. The team began with a simple rule - requiring specifications for tasks above three story points - and focused on one core skill: writing specs. These specs, treated as critical as source code, evolved into two types: frozen (locked after completion) and live (continuously updated). AI played a key role in generating and refining specs, reducing ambiguity and improving async collaboration between product and engineering teams. An AI guild was formed to iteratively improve practices, ensuring continuous learning and adaptation.

AI's broader impact on software development is explored, including accelerated delivery cycles - some features shipped in under 24 hours - and a shift in engineers' roles from writing code to reviewing AI-generated output. This has led to increased merge frequency and delivery rates, especially among early AI adopters, but also introduced challenges like information overload and the need for better context management. The team emphasizes measuring AI's impact across different user cohorts and rethinking traditional KPIs. Looking ahead, AI agents are expected to automate inner development loops (coding, testing, reviews), while humans focus on higher-level decision-making, moving toward a "software factory" model that prioritizes efficiency, feedback loops, and business alignment.

• What if you implemented a spec-driven workflow for your next feature using AI assistance?

  • Move: Before writing any code, use an AI tool (e.g., GitHub Copilot or Cursor) to generate a structured spec document based on a user story - include goals, acceptance criteria, edge cases, and integration points. Treat this as your "frozen spec" and only proceed to coding once it's finalized.
  • Why Now?: AI can now produce reliable, context-aware specs in minutes, reducing ambiguity and rework; early adopters are already seeing 2 - 3x faster delivery cycles.
  • Expected Upside: Reduces back-and-forth during development, aligns stakeholders asynchronously, and cuts debugging time by ensuring clarity upfront - especially valuable when working solo and needing to simulate team rigor.

• What if you formed a personal AI guild to refine your solo development practices every two weeks?

  • Move: Set a recurring biweekly 60-minute session with yourself (block calendar) to review your last two sprints: assess which AI-assisted tasks succeeded, where specs failed, and update your personal "skill playbook" (e.g., how you prompt, validate, and version specs).
  • Why Now?: The pace of AI tooling evolution (every 3 - 6 months) means stagnation risks inefficiency; structured reflection mimics high-performing teams and prevents drift.
  • Expected Upside: Builds a feedback loop into your workflow, turning isolated tasks into cumulative skill growth - leading to more consistent output, fewer bugs, and faster iteration over time.

• What if you made AI-generated specs mandatory for all tasks above 3 story points - even when working alone?

  • Move: Define a personal rule: any task estimated at >3 story points (or >3 hours) requires a live spec document co-authored with AI, updated throughout development, and archived upon completion. Use it to track decisions, edge cases, and integration logic.
  • Why Now?: As code volume grows via AI, mental context becomes a bottleneck - specs act as externalized memory and prevent oversight in complex solo projects.
  • Expected Upside: Enables scalable self-collaboration across time (e.g., resuming work after breaks), reduces regression risks, and creates reusable knowledge assets that accelerate future features.

• Implement a mandatory spec-writing process for all development tasks exceeding 3 story points, using AI-assisted tools to generate initial drafts and ensure alignment with organizational standards.
• Create a bi-weekly AI guild or peer forum to review spec quality, share best practices, and iteratively improve AI-driven development workflows based on real project feedback.
• Start with a single high-impact skill - such as AI-generated specification writing - and standardize it across your workflow before expanding to other AI-assisted skills like test planning or documentation.
• Integrate frozen and live specs into your development cycle: use frozen specs as auditable, post-completion records and maintain live specs for actively evolving features to ensure clarity and traceability.
• Measure AI adoption impact by tracking granular KPIs (e.g., merge frequency, delivery speed) segmented by user cohorts (e.g., heavy vs. light AI users) to identify bottlenecks and optimize tooling for sustainable productivity gains.