How I AI: Ryan Nystrom's 3 Notion Workflows for Engineering Velocity

In this episode of How I AI, I was so excited to sit down with Ryan Nystrom from Notion. As an engineering manager who still loves to write code, Ryan is living at the intersection of team leadership and hands-on building, and he’s using AI to excel at both. He shared some incredible workflows that are not only making his team more productive but also making work more fun and creative.

For anyone who has been in engineering for a while, Ryan's story is familiar. He'd worked the same way for over a decade, but in the last year, everything has changed—his IDE, his terminal, his tools. But instead of being a source of stress, this constant change has brought a new energy and joy to his work. He’s part of a team at Notion that is notoriously "very, very AI pilled," and they’re bringing that puppy dog energy to some of the hardest problems, like developer experience.

We looked at three specific, powerful workflows that Ryan and his team use every day. First, how an AI agent completely automates his standup prep, turning what was once a chore into the most productive part of the day. Second, we saw how his team can ship a feature from a simple text message to a full pull request in 20 minutes, just by @mentioning an agent in a Notion comment. And finally, we explored a spec-first development process where engineers write plain-English documents and an AI agent handles the implementation and verification.

These aren't just theoretical ideas; they're practical, implemented systems that are changing how software is built at Notion. They show a future where engineers are freed from tedious, repetitive tasks to focus on what they do best: architecture, systems thinking, and solving hard problems. Let's get into how he does it.

Workflow 1: Never Prep for a Standup Again with a Custom Notion Agent

Every engineering manager knows the difference between a good standup and a bad one. The bad ones are a dead-eyed slog where everyone just reports what they did yesterday. The good ones are dynamic, focused on problem-solving, and a source of energy for the team. The challenge is that preparing for a good standup—gathering context, pulling metrics, and setting an agenda—is a time-consuming, soul-sucking task. Ryan's solution? He automated the whole thing.

His team has a custom Notion AI agent, affectionately nicknamed "Hot Potato," that acts as a pre-meeting assistant. It allows Ryan to code right up until the minute the meeting starts, confident that a comprehensive, detailed pre-read is waiting for the team.

Here's the step-by-step process:

1. Automated Trigger: A recurring Notion automation creates a new, blank meeting page from a template every morning before the standup.
2. The Agent Wakes Up: At 9 AM every day, the "Hot Potato" Custom Agent is triggered. Its first instruction is to look back over the last 24 hours of activity.
3. Gathering Context: This is where the magic happens. The agent fans out and pulls information from multiple sources simultaneously:

• Metrics: It uses a Honeycomb MCP (Metrics, Content, and People) integration to pull the latest CI performance metrics.
• Communication: It scans the team's project channel in Slack for updates, feedback, and open questions.
• Work: It looks at the team's task database in Notion for any recently closed tickets.
• Code: It checks for merged pull requests in GitHub.
• History: It even reads the previous day's meeting notes to ensure continuity.

1. Synthesizing the Pre-Read: The agent then compiles all this information into a structured document using a template Ryan provided in the agent's instructions. The format is clear and actionable:

• CI Speed
• Decisions
• Progress & Changes
• Bugs & Feedback
• Open Questions & Risks

1. Notifying the Team: Once the page is populated, the agent posts a "brief and fun" message in the team's Slack channel with a link to the pre-read. The meeting starts with everyone already on the same page, ready to discuss what truly matters.

This workflow saves Ryan about 20 minutes of manual prep time each day, but more importantly, it protects his focus and energy for creative work. It also democratizes information, ensuring that the quiet but brilliant engineer's work is highlighted just as much as anyone else's.

Workflow 2: From Text Message to Pull Request in 20 Minutes

This workflow is a perfect example of reducing friction to an absolute minimum. What if you could get a bug fix or a small feature implemented just by writing it down and asking an agent to do it? That's exactly what Ryan's team has built with an internal system they call "Boxy."

The story started with a text from a friend. The friend liked Notion's tab block but wished he could copy a direct link to a specific tab. Ryan thought it was a great idea. In the old world, this would mean creating a ticket, prioritizing it, assigning it, and eventually, someone would get to it. In Ryan's world, it's much faster.

The "Boxy" Workflow:

1. Create a Clear Task: Ryan created a task in Notion. He wrote a few sentences describing the feature, noted a related hover-state bug he saw, and even anticipated an edge case for deep linking. He dropped in a screenshot to show exactly where the new "Copy link to tab" button should appear.
2. Invoke the Agent: In the comments of that Notion task, he simply mentioned the coding agent: @Codex.
3. The Agent Gets to Work: This @mention is the trigger. It kicks off "Boxy," Notion's internal background agent system. A virtual machine is spun up, complete with Codex, the full codebase, and all the necessary developer tools. The agent reads the task description, the comments, and even analyzes the screenshot for context.
4. Review the AI-Generated PR: In about 20 minutes, the agent commented back on the original Notion task with two links: one to a completed pull request on GitHub and another to a live preview environment. The PR was incredibly detailed, including not just the code but also a description of how the agent tested its own work, complete with screenshots of the UI verification it performed.

When Ryan found a line of code he didn't understand, he didn't have to sheepishly debug it. He just commented on the PR, "I don't know what is going on here," and the agent explained its reasoning and fixed the remaining type errors. This isn't just about speed; it's about a new, more fluid, and less emotionally fraught way of collaborating on code.

Workflow 3: Spec-First Development Where the Agent Writes the Code

This final workflow represents a fundamental shift in the role of a software engineer. Instead of spending time on the plumbing and implementation details, engineers can focus on architecture and behavior. The core idea is to write a detailed, human-readable specification for a feature and then let an AI agent do the hard work of writing, testing, and verifying the code.

When rebuilding the harness for Notion AI, Ryan's team took inspiration from modern coding agents and decided to start with specs, not code. He walked us through building a feature called "Ask Mode," which restricts the AI to only use read-only tools.

The Spec-First Process:

1. Dictate the Idea: Ryan didn't even type. He just opened Whisper and started talking, explaining how "Ask Mode" should function. This raw brain dump became the source material.
2. Generate a Formal Spec: He then fed this raw transcript to Codex with a simple prompt and a few examples of existing specs from their library.

"Here's our other like spec library. Learn the format, take my information, write a spec."

1. Implement from the Spec: After a few revisions, Ryan had a comprehensive markdown document. He checked this ask_mode.md file into a special /agent-specs/ folder in their repo. Then came the next step. He pointed Codex at the spec file and gave it a simple command:

"Build it."

1. Autonomous Build & Verify: This is where the process becomes truly powerful. The agent didn't just write the code. It used a detailed Verification section within the spec and a custom CLI tool to continuously test its own work. It would spin up Notion AI, send it queries, enable and disable "Ask Mode," and check the outputs against the behavior defined in the spec.
2. The Spec is the Source of Truth: After a couple of hours, the feature was done. Now, when the team wants to update or change the feature, they don't touch the code. They update the Markdown spec file. The spec is the version-controlled source of truth for how the feature works, making it an invaluable, human-readable asset for engineers, product managers, and even marketing to understand.

The Future is Fast and Fun

What I love about these three workflows is how they all point to the same conclusions. First, developer experience is no longer a nice-to-have. In the age of AI agents, a slow CI/CD pipeline is a mathematical limit on your team's output. As we learned from Steve Kaliski at Stripe, you can't have agents shipping 1,300 PRs a week if they're waiting an hour for builds to complete. Speed is paramount.

Second, the role of the engineer is elevating. We're moving from being implementers to being architects, systems thinkers, and teachers. Our job is to create the specifications, the verification loops, and the tools that enable agents to do the tedious work, freeing us up for more creative, high-level problem-solving.

Finally, this new way of working is just more fun. Ryan said it best: he's having more fun, getting more done, and feeling more energized than ever. We used to be told we could have our projects be good, fast, or cheap—pick two. With AI, it feels like we can finally get the whole triangle. By automating the toil, we're not just gaining productivity; we're rediscovering the joy in building.

Thanks to our sponsors!

• WorkOS—Make your app Enterprise Ready today
• Orkes—The enterprise platform for reliable applications and agentic workflows

Episode Links

• YouTube: https://youtu.be/pUHA_jNwuYE
• Spotify: https://open.spotify.com/episode/7Lm9bBEKovn2uBAihXW5zv?si=6qYqpVR5Tve2mUtqBP5GPQ
• Apple Podcasts: https://podcasts.apple.com/us/podcast/spec-driven-development-the-ai-engineering-workflow/id1809663079?i=1000767179882