{ "video": { "id": "1et3Lb_KLQo", "title": "I Don't Fear AI Slop. I Fear Untold Stories - Zack Proser, WorkOS", "duration": 1114, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "" }, "analysis": { "video_id": "1et3Lb_KLQo", "title": "I Don't Fear AI Slop. I Fear Untold Stories - Zack Proser, WorkOS", "one_liner": "Zack Proser argues the bigger risk isn't AI slop but the stories that never get told. If you've made a type of artifact long enough to have strong opinions about what makes it good, put those opinions into an automated \"factory\" and give it to people who don't normally create, while a human still inspects and signs off on every piece.", "summary": "Proser has written obsessively for years; he once wrote and published his best-performing post in 23 minutes, before LLMs. He says he worries less about AI slop, which he thinks can be fixed with technique, than about talented colleagues who never share their work because of fear, shame or a belief that they're 'not a writer'. At WorkOS he built Blog Bot, a Slack-fronted system on Cloudflare that contains his writing taste and produces drafts that are 85-90% done. He uses a cymbal production line and a blacksmithing metaphor to argue that craft survives high speed as long as a human shapes the work and signs off on it. He then repeated the pattern for an audience of two: a bedtime-story app for his toddlers, built in 4 hours.", "key_points": [ "Before LLMs, Proser wrote and published his consistently best-performing post, about 15 years of running a self-published tech blog, in 23 minutes after a hard bedtime with his two toddlers. His explanation for writing fast is that he is 'always writing'.", "Blog Bot runs on Cloudflare, and colleagues use it through Slack (e.g. '@Blog Bot, here's a transcript from an event, produce an on-brand WorkOS recap post'). It returns a draft 85-90% of the way there, which a human must read all the way through, iterate on, put their name on and publish.", "'Asymmetric leverage': a very simple interface hides a complex distributed system with data connectors to every company information source (Granola recordings, event images, Slack discussion, planned framing), custom tools, human approval to confirm facts, and an image-generation endpoint whose outputs match the WorkOS design team's style.", "Colleagues can adjust and tune their drafts without Proser or any other writer being involved. Publishing rates rose across the org, and colleagues who did excellent work but rarely shared it began publishing regularly.", "WorkOS founder Michael Grinich showed him a video of cymbals being mass-produced, with a long-serving musician playing and signing off on each one, and said 'This screams craft to me.' His point was that craft runs through the entire line even when machines do the heavy lifting.", "Blacksmithing metaphor for writing with an LLM: he is the spark (taste: deciding a story is worth telling), the bellows (continuously shaping and cutting while the LLM does the tedious work) and the quench (signing off: 'Zack Proser said this and he stands behind all these words'). He reaches the end goal about 6x faster with many more iterations, e.g. 14 iterations over 2.5-3 hours.", "If the end user can't tell hand-writing from LLM-iterated work, the difference doesn't matter. The real rubric: is it true? Is it useful? Is it beautiful? Does it help anybody else?", "His second factory, for exactly two people: a bedtime-story app built with Opus 5.5 extra high in 4 hours. It reads photos of his kids' day, turns them into a narrated, animated story that casts the kids, adds real-time voice control (saying 'again again again' replays it) and makes custom stories such as 'me and Captain America at the zoo' in about a minute and a half. On average each story gets about three repeat requests, his 'most successful playtest' ever." ], "takeaways": [ "Find the artifact or workflow you know by heart, the thing colleagues say you're 'so good at', and codify your strongly held opinions about what makes it good into an automated tool.", "Put the factory behind an interface people already use every day (e.g. Slack), and have it gather the context automatically so users don't have to collect it by hand.", "Keep a human accountable at the end: drafts aren't one-shot to production, and someone must read the whole thing and put their name on it. Judge output on whether it's true, useful, beautiful and helpful, not on whether AI touched it.", "Hand the factory to people who don't normally produce that kind of work, let them tune it without you, and improve it based on their feedback.", "Consider very small audiences: with LLM-accelerated development, even two people can be worth building a factory for. If you're not a developer, have an agent help you build it." ], "topics": [ "ai writing", "content automation", "developer relations", "craft and quality", "internal tools", "agents", "storytelling", "slop" ], "tools": [ "WorkOS", "Blog Bot", "Cloudflare", "Slack", "Granola", "Opus 5.5", "YouTube" ], "quotes": [ { "text": "I'm significantly more concerned about all the stories that go untold.", "at": "03:13", "url": "https://www.youtube.com/watch?v=1et3Lb_KLQo&t=193s" }, { "text": "This screams craft to me.", "at": "08:05", "url": "https://www.youtube.com/watch?v=1et3Lb_KLQo&t=485s" }, { "text": "the real rubric here is is it true? Is it useful? Is it beautiful? And does it help anybody else?", "at": "12:53", "url": "https://www.youtube.com/watch?v=1et3Lb_KLQo&t=773s" }, { "text": "who will you build a factory of loving grace for?", "at": "17:42", "url": "https://www.youtube.com/watch?v=1et3Lb_KLQo&t=1062s" } ], "words": 4549 }, "summary_url": "/#1et3Lb_KLQo", "transcript": { "html": "/transcripts/1et3Lb_KLQo.html", "txt": "/transcripts/1et3Lb_KLQo.txt", "vtt": "/transcripts/1et3Lb_KLQo.vtt" } }