Kieran Klaassen demonstrates the AI software factory he built for Cora, Every's AI email app. Teammates submit feedback in Slack, a scheduled Claude Cowork routine analyzes it and creates a structured pull request, and Fable 5 with Compound Engineering's LFG workflow implements the fixes in a single batch. The walkthrough covers turning Slack into a feedback inbox, capturing feedback with RiffRec, automatically converting feedback into a pull request, processing 17 pieces of feedback in one batch, reviewing the changes through video walkthroughs, teaching the system from its mistakes, and letting it merge working improvements autonomously. # Channel Identity Observations — Video: "How I Built an AI Software Factory With Fable 5" ## Explicit Self-Description Evidence - **No formal channel description provided** — the video contains no intro slogan, tagline, or explicit statement of what the channel is about. Identity must be inferred from presentation only. - **Personal sharing stance**: the creator explicitly frames the content as anecdotal, not prescriptive: > "I'm just sharing what I do now. I'm not saying this is the way, but this works for me" - **Audience invitation at the end** suggests a participatory community: > "please let me know, uh, what you think, how you improve this" ## Observable Presentation Patterns ### Intended Audience - **AI-assisted software developers and builders** — the content assumes familiarity with tools and concepts like **Cursor**, **Slack MCP**, **pull requests**, **CI checks**, and **LFG flow**, none of which are explained for beginners. - **Practitioners of the creator's own ecosystem**: he references **"Compound Engineering"** (a method/tooling he points viewers to: "go check it out") and an internal alpha-feedback channel with named colleagues (e.g., "Brandon is dropping all kinds of information"), suggesting an audience overlapping with **Every** — the company where "people use this internally at Every and myself." ### Claimed / Demonstrated Scope - **Building AI-driven development workflows**: the video demonstrates a "software factory" pipeline — scheduled Slack feedback ingestion (via "Clockwork"), classification, automated PR generation, batch AI fixing, and autonomous merge. - **Product building in public**: rebuilding "Cora V2," described as "a full inbox," with an alpha channel of internal testers. - **Open-source contribution**: creator states, "Riffraff is a library, it's open source that I created" — a wrapper recording DOM clicks, voice, network requests, and errors for feedback capture. - **AI model enthusiasm/evaluation**: opens with strong endorsement of "Fable": "I've been using Fable for about a week now and I love it. It's my favorite model ever." ### Presentation Style (Demonstrated) - **Screencast walkthrough format**: shares screen showing Slack channels, scheduled tasks, YAML files, PR branches, and video "walk-throughs" of AI fixes. - **Conversational, unpolished delivery**: heavy filler ("um," "uh"), first-person narration, spontaneous reactions ("this is the first time I look at this"). - **Emphasis on speed and autonomy as value props**: "how insane the speed is that I can now ship features with"; the "magic moment" of overnight autonomous merging while sleeping. ## Claims vs. Demonstrated Distinction - **Claim (creator statement)**: the workflow "works for me"; batch review is "manageable"; autonomy "feels like a superpower." - **Demonstrated**: screen recordings of the Slack alpha channel, the scheduled task output, PR contents, and AI-generated fix videos — the pipeline itself is shown, not merely described. - **Claim**: Riffraff is open source and created by the speaker. - **Not verifiable from material**: audience size, channel frequency, whether this workflow generalizes beyond Cora/Every. # Delivery Observation: "How I Built an AI Software Factory With Fable 5" ## Overall Tone - **Enthusiastic and informal**: Opens with direct praise — *"I've been using Fable for about a week now and I love it. It's my favorite model ever."* - **Conversational, stream-of-consciousness style**: Frequent self-corrections and restarts, e.g. *"So it's it's basically a factory where lots of stuff goes automatic."* - **Humility framing**: Repeatedly hedges his approach as personal, not prescriptive — *"I'm not saying this is the way, but this works for me."* ## Vocabulary & Recurring Phrases - **"So" as a primary connector**: Dozens of utterances begin with "So" — *"So what you're looking at is…"*, *"So, you can see this skill here."*, *"So, yeah."* - **Repeated filler sounds**: *"Um"*, *"uh"*, and *"yeah"* appear throughout, giving an unscripted demo feel. - **Superlative/enthusiasm vocabulary**: *"it feels like a superpower"*, *"the magic moment was"*, *"how insane the speed is"*, *"which is really cool."* - **Repeated intensifier "kind of"**: *"which is kind of cool"*, *"it's kind of interesting"*, *"you can kind of uh, look at your app."* ## Teaching Habits - **Show-then-explain pattern**: Narrates on-screen actions as they happen — *"You can see screenshots. You can see riff rack zips. I'll share you what that is."* - **Question-and-answer asides**: Anticipates viewer questions — *"Um what is riffraff? Riffraff is a library, it's open source that I created."* - **Deferred explanation**: Flags topics to return to — *"I'll share you what that is"*, *"We'll get into that next."* - **Live, unedited walkthrough**: Admits on-camera discovery — *"this is the first time I look at this. I've not looked at uh, all of the things."* - **Explicit learning summaries**: *"but the learning here is that I mine the stuff in batches and I kind of like that."* ## Communication Patterns - **Honest caveats about limitations**: Acknowledges manual steps and costs — *"what I do still manually, but I could automate this"*, *"It takes a long time to run. Sometimes like two, three, four hours."* - **Storytelling payoff structure**: Builds to an anecdote — *"this morning I woke up and Brandon said, 'Woah, like, the design looks good.'… I was sleeping and it was doing that. So, that's the dream."* - **Calls to action at close**: Invites feedback — *"please let me know, uh, what you think, how you improve this."* - **Cross-promotion habit**: References his own ecosystem repeatedly — Compound Engineering, Riffraff, Cora — *"go check it out. Uh, it's in Cursor, it's wherever you want."* ## Uncertainty Notes - Sample is a **single video**, so patterns like the "show-then-explain" structure and heavy "So"/"um" usage cannot be confirmed as stable across the creator's body of work. - The enthusiasm and personal-anecdote style appear consistent throughout *this* video but may vary with topic. # Position Record: "How I Built an AI Software Factory With Fable 5" ## Speaker - Single, unnamed narrator (no curated speaker metadata provided); identified only as the video creator, an internal builder at "Every." ## Overall Stance on Fable (the model) - **Strong enthusiasm, framed as personal experience:** "It's my favorite model ever. It feels like a big shift." - Claim: "The stuff that I wanted to work suddenly started working." - Calls the resulting workflow "a superpower." ## Core Project: Cora V2 - Rebuilding **Cora computer** as version two, described as "a full inbox." - Users: internal team at Every plus the narrator; feedback shared via a **Cora V2 alpha channel** in Slack. ## Stated Workflow ("AI Software Factory") ### 1. Feedback collection - Internal users share feedback, screenshots, and "riffraff" zips in Slack. - **Riffraff** (narrator's open-source library): a wrapper around a React app that records **DOM clicks, audio narration, network requests, and errors**; saves a shareable file to Slack. - Claim: it's "rich information, better than a video," though video also works. ### 2. Automated triage - A scheduled **Clockwork** routine runs morning and evening (sometimes manually). - Uses a skill ("alpha feedback pulse") to: - Fetch Slack messages via Slack MCP - Classify feedback - Check if items are already resolved - Download recordings for analysis with "compound engineering" - Output: structured **YAML tracking** and markdown files of what works/doesn't; a **pull request** is created. ### 3. Batch fixing (semi-manual) - Narrator runs the "LFG flow" (from **Compound Engineering**, usable in Cursor) on the combined PR branch, twice daily. - Instructs it to: fix items, leave notes where human input is needed, consult a **strategy doc**, and make decisions where it can. - Produces a **video walkthrough** and screenshots of fixes. ### 4. Human review - Narrator checks out the branch, reviews videos, and uses `/polish` for a side-by-side server preview on larger changes. ## Key Reasoning and Claims - **Batching rationale:** reviewing one combined PR with ~17 items is far more manageable than 17 separate pull requests. - **Self-correction claim:** the "Compound step" means a mistake "will also not make the same mistake the next time." - **Speed claim:** runs take "two, three, four hours." - **Magic moment (anecdote):** instructed the system to merge if everything looked good and CI was green; woke up to positive feedback from Brandon ("the design looks good") — "I was sleeping and it was doing that. So, that's the dream." ## Qualifications and Caveats (preserved as stated) - "I'm just sharing what I do now. **I'm not saying this is the way**, but this works for me." - The review step is **still manual** — narrator admits "I should probably automate this." - Acknowledges incomplete oversight: "I don't even know exactly everything that was reported... this is the first time I look at this." - Runtime cost: multi-hour execution. ## Closing Position - Invites viewer input: "Please let me know what you think, how you improve this." - Plans to incorporate these learnings into **Compound Engineering** and make it "a little bit more robust"; still "trying out new tools." # Speaker Attribution Record — "How I Built an AI Software Factory With Fable 5" ## Video Metadata - **Video ID:** `01a0f15a-82b1-737f-a783-47e5a36e1acc` - **Title:** How I Built an AI Software Factory With Fable 5 - **Curated metadata:** Not supplied ## Speakers & Roles | Speaker | Role | Evidence | |---|---|---| | Unnamed first-person speaker ("I", "Hello everyone") | **Creator** | Narrates personal workflow, project ownership ("my beautiful project Cora"), and direct address to audience | - Only one speaking role present: the **creator**, throughout the entire transcript. - **No guests, narrators, or other identified speakers** appear. ### Third Parties Referenced (non-speaking) - **Brandon** — mentioned twice, not speaking: drops feedback in the Cora V2 alpha channel; later quoted indirectly: *"Woah, like, the design looks good"* / *"Things are slick."* - Unnamed **internal users at Every** — referenced as testers providing feedback. ## Provenance & Referenced Sources ### Tools / Products Cited by Creator - **Fable (5)** — AI model; creator states he has used it "about a week." - **Cora / Cora V2** — creator's own project ("I'm rebuilding a version two of Cora which will be a full inbox"); alpha channel in use at **Every**. - **Clockwork** — hosts the creator's scheduled routine (morning/evening runs). - **Slack / Slack MCP** — feedback channel and message-fetching mechanism. - **Riffraff** — open-source library the creator says he created; React wrapper recording DOM clicks, audio, network requests, and errors. - **Cursor** — IDE used with the workflow. - **Compound Engineering** — referenced as the source of the "LFG flow" and the "compound step" (mistake-avoidance); the creator states he may incorporate his learnings into it. - **{slash} polish** — command/tool used for review servers. ### Claims About Third-Party Contributions - **LFG flow:** creator attributes it to **Compound Engineering**: *"the LFG flow is from Compound Engineering. Go check it out."* - **Riffraff:** creator claims authorship: *"it's open source that I created."* ## Promotional Segments - **No explicit promotional segments** (no sponsor reads, ads, or affiliate promotions in the transcript). - Note: the creator offers repeated positive endorsement of Fable (e.g., *"It's my favorite model ever"*) and directs viewers to Compound Engineering — endorsement, not flagged promotion in the source. ## Comment Availability - **Not stated.** No curated metadata supplied; transcript invites viewer response (*"please let me know what you think"*) but does not establish comment availability. ## Uncertainty Notes - Speaker name not identified in transcript or metadata. - "Fable 5" appears only in the title; the transcript refers to the model only as "Fable." - Role attributions (creator) inferred from first-person ownership statements; no conflicting evidence present. Hello everyone. I want to do a video. I've been using Fable for about a week now and I love it. It's my favorite model ever. It feels like a big shift. The stuff that I wanted to work suddenly started working. I want to show you what I have now and how it works. So what you're looking at is my beautiful project Cora computer I'm rebuilding a version two of Cora which will be a full inbox. So you can use it like that. And yeah, like I'm building that right now. I just want to show you how I do that and how insane the speed is that I can now ship features with and what this look look like. So it's it's basically a factory where lots of stuff goes automatic and yeah, I'm just sharing what I do now. I'm not saying this is the way, but this works for me and yeah, it's kind of interesting. So let me just share you what I do. First of all, we have people use this internally at Every and myself and we share feedback. So we have a Cora V2 alpha channel you can see Brandon is dropping all kinds of information. You can see screenshots. You can see riff rack zips. I'll share you what that is. And you can see here that there are these eyes which means I am looking at it. You can also see before here if you scroll a little back, there are some check marks of solved. So this is done. And and how that works is the following. I have a scheduled routine in Clockwork that will run and in the morning and the evening and sometimes I do it manually and the idea is that it goes through Slack. Just reads everything and structures it. So, you can see this skill here. This is the alpha feedback pulse. Basically, it loads everything, it fetches messages using the Slack MCP, classifies it, checks if things are already done, if there are recordings like riffraff recordings or video, it downloads that so we can analyze them with compound engineering. Um it will say use the skill to analyze them. So, there's a bunch of stuff in here. Um what is riffraff? Riffraff is a library, it's open source that I created. It's just a wrapper that you can wrap around your React app and basically you can have a button like this where you click and say, "Okay, let's allow." And now it's recording feedback. It records what DOM you're clicking on, it records what you're saying, it records all the network requests and errors. And if you hit stop, it will save this file and you can share that to Slack. That's how we do it at least. So, it's like rich information, better than a video. But you can do a video as well, it works with that as well. So, just get people to share feedback. That's the the first thing. And you can do it yourself as well. The next this will be processed. So, the scheduled task you can see this one for example is Um yeah, it just goes through it like, "Hey, reading, downloading all the files, processing." And I can pull request is created and you can see here um this is a YAML file how it structures it. So, it keeps track of what it reported, what it didn't. There are markdown files of the things that work and don't work that it figured out here. So, this is the information. Then, what I do still manually, but I could automate this. I should probably automate this, but what I do then now there's this in the morning and in the evening, so twice. So, all this together, I just go to uh cursor and I use the cloud and say LFG flow for each of the every items in the pull request. Let's work in this PR branch, too. So, like I combine it, which is what I like here. Fix all of the items, leave notes for the ones that need human input, but look at the strategy, makes a decision if you think you can make a decision. So, there's a strategy doc. And basically, it's uh well It's loading everything here, but it just fixed everything. And you can see here that uh in the walk-through, it just uh shows you um in video form all of the things it fixed. Um which is really cool. So, you can see here like it's changing. Like, I don't even know exactly everything that was reported, which is kind of cool. Uh we'll get into that next, but this is the first time I look at this. I've not looked at uh all of the things. You can see screenshots as well. So, yeah, it's it's there and what I do then is I um uh check out the branch, take the videos. If it's big, I use \{slash\} polish, um which is like giving you a server to look at and uh you can kind of uh, look at your app on the right and on the left. Um, so then I go through and check, um, if everything looks good. Um, and the cool part is the LFG flow is from Compound Engineering. Uh, go check it out. Uh, it's in Cursor, it's wherever you want. Uh, but the learning here is that I mine the stuff in batches and I kind of like that, uh, because it's manageable for me to review. Just imagine like there like 17 items, I think, that were found in Slack. Like, imagine you have to have to review 17 pull requests. It's It's a lot of work. Um, and with Fable, it's very good at doing that. Uh, and yeah, and if if you have more feedback after this, like, it will come up there again and it will be refined. And the cool part is there's the Compound step in Compound Engineering. So, if it makes a mistake, it will also not make the same mistake the next time. Um, so, yeah. Um, that's how I do it now and it feels like a superpower. It takes a long time to run. Sometimes like two, three, four hours, but the magic moment was, uh, last night where I kicked it off and I just said, "Okay, if everything looks good, uh, and the CI is green, merge it." And this morning I woke up and Brandon said, "Woah, like, the design looks good." Or like, "Things Things are slick." And I like, "Oh, that's cool." I was sleeping and it was doing that. So, that's the dream. Um, uh, please let me know, uh, what you think, how you improve this. Uh, I'm going to see if I can incorporate these things in Compound Engineering as well. Um, yeah, make it a little bit more, uh, robust, but I love it so far. Uh, trying out new tools, as well.