Shopify's AI Agent Works in Public Slack. That Might Be the Best AI Training Program in Tech.

The best AI training at Shopify isn't a mandatory module or a Thursday workshop. It's a Slack channel where anyone can watch what the AI does, how colleagues phrase their requests, and what gets revised before it gets used. The training happens because the work is visible — not because the training was planned.

What Shopify built

Tobias Lutke, Shopify's CEO, described this in a post covered by Simon Willison on May 11. River — Shopify's internal AI coding agent — operates only in public Slack channels, never direct messages. Lutke's name for this setup is Lehrwerkstatt: a German term for a teaching workshop where apprentices learn by watching craftspeople work, not by sitting through instruction.

Willison drew a direct parallel to Midjourney's early days on Discord, where watching other users generate images taught you what worked faster than any tutorial. The mechanism is the same: proximity to someone else's process is a more efficient teacher than documentation.

The structural point is simple. When AI work happens in DMs or private browser tabs, the only person who learns anything is the person running the session. When it happens in public, every search, every "how did you get that output?" question becomes a teaching moment without anyone scheduling one.

What to do Monday morning

Most teams won't build a custom agent. But the principle doesn't require one.

The minimum version: pick a Slack channel — or a Teams thread, a Notion page, wherever your team already talks — and agree that some AI interactions get posted there. Not polished outputs. The actual exchange: the prompt, the response, what got revised, what was kept. You want the process visible, not just the result.

The rule Shopify uses — public by default, never DMs — is stricter than most teams need. Start with "post the ones that taught you something." That's enough to get the channel working.

A content team that tried it

A content team lead at a 55-person B2B software company started a Slack channel called #ai-prompts-and-outputs in January. The rule: if you use AI to produce something that goes to a client or gets published, paste the key exchange — prompt, response, and any follow-up.

Three months later, the team's prompts had shifted without a single training session. People stopped writing vague single-sentence instructions after watching a colleague's detailed briefing generate usable copy in one pass. A new writer asked why one post used a specific structural approach. The answer was already in the channel, searchable. Nobody had to write a guide.

The limits of this model

A solo founder running all AI work alone gets nothing from visible channels — there's no one watching. And an organization where sharing work is politically charged can make the channel a source of anxiety rather than learning.

Willison's Midjourney comparison is useful here: Discord worked because generating images was low-stakes and high-delight. The more evaluative your culture, the more you need to frame the channel explicitly as a learning record rather than a quality review.

The setup that tends to work: the channel exists to share what taught you something, not to showcase clean outputs. Rough prompts belong there as much as polished ones. An ops manager at a 30-person logistics company told her team: "Post the failures too. Especially those."

The takeaway

Shopify spent nothing on training River's users. They just made the work impossible to miss.