Merge rate jumped from 36% to 77% in two months. No model change. No retraining. The only thing Shopify did was block their AI agent from receiving direct messages.
What's This AI Agent That Refuses DMs?
Shopify has an internal AI coding agent called River. It reads code, runs tests, opens pull requests, queries data warehouses, and inspects production traces. But it has one unusual policy — it won't respond to DMs.
Send it a direct message and it politely declines, suggesting you create a public channel to work in together. It sounds like an odd constraint, but it's deliberate. Shopify calls this a Lehrwerkstatt — a "teaching workshop" in German.
The idea is simple. Just like an apprentice in a workshop learns by sitting next to a master craftsman, when all AI work happens in a visible space, learning happens without any formal curriculum. Simon Willison immediately recognized this as the same principle behind Midjourney's Discord strategy. When Midjourney's early users had to generate images in public Discord channels, they naturally absorbed AI prompting techniques by watching each other. It's one of the secrets behind how a team of fewer than 50 people built a $200M+ ARR business.
"The whole shop floor is the classroom. You learn by being near the work."
— Tobias Lütke, Shopify CEO
How Is This Different From Private AI Use?
Why does AI adoption stall in most organizations? Even with a ChatGPT Enterprise license, half your team uses it in private windows — and the other half doesn't use it at all. Great prompts stay locked in one person's head. Failures get buried. Nothing spreads.
The public channel model flips this entirely. River's merge rate didn't jump from 36% to 77% because the model improved. It happened because employees shared River's failure cases in public channels, and teams documented improvements as "Skills" modules — collective learning in action. Tobi Lütke himself runs #tobi-working-with-river, where 100+ employees watch him work with AI in real time.
| Private AI Use (Default) | Public Channel AI (River Method) | |
|---|---|---|
| How learning happens | Individual trial and error | Osmosis — absorb by watching |
| Failure handling | Buried individually | Shared → collective improvement |
| New hire onboarding | Start from scratch | Channel history becomes training material |
| Prompt sharing | Rarely spreads | Propagates organically across the org |
| AI skill gap | Individual gaps widen over time | Average rises together |
Josh Bersin Research
Companies that achieve AI-native "dynamic knowledge sharing" are 6x more likely to exceed financial targets. The core of organizational learning isn't content — it's visibility.
The Essentials: How to Apply This to Your Team
- Create a dedicated public AI channel
Set up a public AI channel in Slack or Teams. Simply declaring "no DMs for AI work" as a team rule is enough to start. Name it something concrete like #ai-work or #ai-experiments. - Let leadership go first
Like Tobi Lütke, have a leader create their own #leader-name-ai-work channel and share prompts, results, and failures publicly. Without visible role modeling at the top, your team will default to private windows. - Run a failure log channel
Create a separate "Things AI got wrong this week" channel. Betterworks ran biweekly "lab meetings" to share failures, preventing teams from rediscovering the same lessons. Sharing failures also builds psychological safety. - Build a shared Skills doc
Collect your team's best prompt patterns in a collaboratively edited doc. Shopify turned this into "Skills" modules applied across River. A thread for "this prompt worked great" will fill it naturally. - Use channel archives for new hire onboarding
Tell new teammates: "Read this channel your first week." Accumulated public work becomes training material. You won't need to formally teach anyone how to use AI.
Dig Deeper
Learning on the Shop Floor Simon Willison's original essay. Shopify River analyzed alongside the Midjourney Discord parallel. simonwillison.net
Inside Shopify's AI-First Engineering Playbook BVP's deep-dive interview with Shopify VP of Engineering. 20% productivity gains, merge rate stats, and the "comprehension debt" risk. bvp.com
Make the Work Visible Real visibility tactics from CarGurus, Fandom, Betterworks, and Red Hat. Four practitioner approaches with outcomes. airealizednow.substack.com
From Memo to Movement: Shopify's Cultural Adoption of AI First Round Capital's analysis of Shopify's AI culture. Legal alignment, unlimited tool budgets, MCP infrastructure. firstround.com
New Research: How AI Transforms $400 Billion Of Corporate Learning Josh Bersin's research on organizational learning. Why dynamic knowledge sharing companies outperform 6x. joshbersin.com
How Midjourney Built a $200M+ AI Business Through Discord-First Community How public channel design drove collective learning and $200M+ ARR with fewer than 50 people. ideaplan.io
