Spec-driven development: The AI engineering workflow at Notion | Ryan Nystrom

What It Covers

Ryan Nystrom, engineering manager at Notion, demonstrates three AI-powered workflows: an automated daily standup agent that pulls from Slack, GitHub, and Honeycomb telemetry; background coding agents triggered by task mentions; and spec-driven development where markdown specifications serve as the source of truth for autonomous code generation.

Key Questions Answered

• •Automated Standup Agent: Build a custom AI agent that runs at 9AM daily, pulling the last 24 hours of Slack messages, merged pull requests, closed tasks, and yesterday's meeting transcript into a structured pre-read. This eliminates meeting prep entirely, letting engineering managers write code until the meeting starts while ensuring every team member's contributions surface equally.
• •Spec-Driven Development Workflow: Start features by voice-dictating requirements into Whisper, feeding the transcript to Codex to generate a structured markdown spec, then pointing Codex at that spec file to build the implementation. Nystrom reports near-complete one-shot results on full features because comprehensive specs include code pointers, edge cases, and explicit verification steps that guide the agent precisely.
• •Background Agent via Task Mention: Configure a VM-based coding agent so engineers can trigger it by mentioning it in a task comment alongside a screenshot or brief description. Nystrom's example: a friend's feature request described in three paragraphs and one screenshot produced a complete pull request with UI verification screenshots in approximately 11 minutes, with CI failures resolved through follow-up comments.
• •CI Speed as Agent Force Multiplier: Reducing CI runtime from slow cycles to under three minutes dramatically increases agent throughput, since agents sit idle waiting for test results. Nystrom's Afterburner project targets cutting CI to one-quarter of its original duration, citing Stripe's 1,300 agent-generated PRs per week as evidence that slow pipelines mathematically cap AI coding capacity.
• •Specs as Living Version-Controlled Documentation: Store feature specs as markdown files inside the repository alongside code. When behavior needs updating, modify the spec first and instruct the agent to implement the delta. The spec history in version control becomes an auditable changelog in plain English, readable by non-engineers for marketing or support purposes without requiring code interpretation.