The Every team spent a week in Panama experimenting freely with the latest AI tools, with no roadmaps or sprints. In this Demo Day session, two team members present what they built: Katie Parrot's Tastemaker, an agent that analyzes any author's work to break down what makes their writing voice distinctive, and Jack Cheng's research assistant, which visualizes complex thinking as interactive visual maps instead of walls of text. # Channel Identity Observations: "What We Built in a Week of Pure Vibe Coding" ## Intended Audience - **Live-streaming audience for an internal-style "demo day."** The speaker opens with, *"Welcome to demo day. So excited to have you. We are coming in live,"* indicating a public or semi-public presentation of otherwise internal work. - **Mixed technical audience.** One presenter explicitly frames her demo for non-technical viewers: *"I think it's interesting for non like I'm deeply non-technical... technical specifics kind of bounce off my head and I get confused."* This suggests the channel addresses both builders and non-builders interested in AI-assisted creation. - **AI-curious creators and writers.** The content centers on AI writing tools, taste, and style guides, e.g., Katie's tool *"helps you analyze your taste"* for people who *"know what they like, but can't explain why."* ## Claimed Scope - **Showcasing experimentation with new AI coding tools.** The host frames the week as exploring how recent model releases change internal building practices: *"we're in one of those moments where software really changed over the last couple months with Opus 45 and Cloud Code and Co-work... to see how that might change a lot of how we build internally at Evergreen."* - **Empowering non-coders to build software.** Katie claims the tooling *"has just like unlocked new capabilities for me even as a non-coder."* - **Educational/community offerings.** The host announces: *"Katie is going to be leading a AI and writing camp very soon. So look out for that,"* indicating the channel promotes courses or camps. ## Observable Presentation Patterns - **Affiliation with an organization ("Evergreen") and editorial team.** Presenters identify roles: Jack, *"a contributing editor"* who wrote *"a piece on AI and creativity,"* and Katie, *"a staff writer"* with a column *"working overtime."* This demonstrates an editorial/publication-backed channel identity. - **Group/retreat-based content format.** The video documents a team "Think Week": *"We have spent the week here on Think Week playing with new technology... We played werewolf last night."* - **Live demo format with named presenters.** The host structures the stream so *"each person who demos, they're going to introduce themselves,"* demonstrating a recurring demo-day presentation style. - **Informal, personality-forward tone.** Casual asides (Werewolf anecdotes, *"Do I get points for saying agent native, Dan?"*) show a conversational, community-flavored presentation. ## Explicit Self-Description vs. Demonstration - **Explicit claims:** The host states the goal is showing how new AI tooling *"might change a lot of how we build internally at Evergreen"* — a direct statement of purpose. - **Demonstrated patterns:** The transcript itself demonstrates hands-on, screen-shared product demos (agent-visualization tools, taste-analysis apps) built with AI coding tools, reinforcing the channel's applied-experimentation focus. No broader channel-level mission statement beyond this video's scope was present. # Delivery Observation: "What We Built in a Week of Pure Vibe Coding" ## Overall Format - **Multi-speaker demo-day structure**: A host frames the event, then each presenter self-introduces and demos. Transitions are handled by the host: > "And next up, we've got Katie." - **Casual, live-streamed register**: Fillers ("um," "uh," "like," "you know") are pervasive throughout all speakers, consistent with unscripted live delivery. ## Tone - **Enthusiastic and informal**: The host opens with high energy and personal color: > "Welcome to demo day. So excited to have you." > "Uh we played werewolf last night. Naveen, it turns out, is incredible at Werewolf." - **Self-deprecating honesty from presenters**: Katie repeatedly flags her own limits: > "I'm deeply non-technical... technical specifics kind of bounce off my head and I get confused." - **Supportive peer reactions**: The host validates each demo: > "like stunning demo. Amazing work. Thanks for sharing, Jack." > "Amazing Katie" ## Recurring Phrases & Verbal Habits - **Filler-heavy restarts and self-corrections**, e.g. Katie: > "it's called, it used to be called Swipe File and now it's called um Taste Maker." > "how much I read by right by hand" - **"You know" as a connective** — appears frequently across speakers: > "you know, they're coordinating coordinating the research" (Jack) > "they know what they like, but they can't explain why" (Katie, no filler but same conversational framing) - **"So" as a transition word** to move between demo steps: > "So let me share my screen." / "So this is the before time version" / "So, this is where the agentic functionality comes in" - **"Um" as a pause-management tool**, especially while navigating live software: > "Um I have the compound engineering plug-in um, plugin set up" ## Teaching / Demo Habits - **Narrate-while-doing**: Presenters describe on-screen actions in real time: > "I'm just going to hit the play button." (Jack) > "Let me switch that." / "let me just search for it" (Katie) - **Problem → solution framing**: Both demos explicitly name a pain point before showing the tool: > "they're doing it in parallel, right? But when we go back and look at transcripts, the transcripts serialize everything" (Jack) > "they know what they like, but they can't explain why" (Katie) - **Before/after comparison**: Katie contrasts the old and new versions explicitly: > "So this is the before time version which is very programmat very programmatic." - **Talking over waiting time**: Katie fills processing delays with humor and jargon callbacks: > "we all sit here and watch it do its thinking and that's very boring. So I'll just say agent native again um to fill the space." - **In-jokes and audience callbacks** (recurring, appears at least twice): > "Do I get points for saying agent native, Dan?" > "I guess I should get Andy Dillard in here to make Dan happy." ## Vocabulary Patterns - **Repeated technical/community jargon**: "agent native" (used 4+ times by Katie, clearly a group term), "compound engineering," "Claude," "spec," "style guide." - **Namedropping of people and publications** as shared-context shorthand: > "Leslie Jameson, Amanda Mole, Gia Tolentino, David Foster, Wallace, you know, the good people." ## Stable Patterns vs. Isolated Moments - **Stable (repeated across speakers)**: filler density, "so"/"you know" connectives, problem-first demo framing, supportive host interjections. - **Isolated / uncertain**: The host's Werewolf anecdote and internet disclaimer occur only in the opening; there is insufficient repetition to call them a pattern. Katie's "Dan" callbacks repeat twice but only within her segment — a habit of *her* delivery, not necessarily the group's. ## Uncertainty Notes - Segment boundaries (`>>`) suggest speaker changes, but the transcript does not label every speaker; some attributions above rely on self-introductions and are otherwise inferred. - Only two full demos are captured; claims about "group-wide" habits rest on thin evidence from the host's brief framing remarks. # Position Extraction: "What We Built in a Week of Pure Vibe Coding" (Demo Day, Part 1) ## Context - Recorded during an **Evergreen "Think Week"** demo day; most participants are together in a house in **Panama**, with some joining remotely. - Opening speaker (unnamed in transcript) frames the week as exploring how recent tooling — **Claude Opus 4.5, Claude Code, and Co-work** — might change internal building practices. --- ## Opening Speaker (unnamed host) - **Claim:** Software development has materially changed in the last few months due to Opus 4.5, Claude Code, and Co-work; the week was an opportunity to test how that changes internal workflows at Evergreen. - Ancillary: Naveen is "incredible" at Werewolf (informal, non-substantive). --- ## Jack (contributing editor, "AI and creativity" piece) **Project:** Visualizer for parallel agent activity. - **Problem claim:** When agents run in parallel, review transcripts serialize that parallel activity, hiding it. A visualization makes parallel work visible. - **How it works:** A real research request's log file drives virtual on-screen characters; events act as triggers for character movement. One lead researcher, four research assistants; agents reconvene to discuss findings and generate a research report. - **Features:** Time-speed scrubbing; follow-a-specific-agent view. - **Stated motivation:** Inspired by what Kieran and Yash were working on; agents "working while we sleep" need better visualization. --- ## Katie (staff writer, "Working Overtime" column; describes herself as "deeply non-technical") **Project:** **Tastemaker** (formerly "Swipe File") — a tool to help people analyze their taste in writing. ### Core position - **Claim:** People "know what they like, but can't explain why." The tool lets users collect clips of writing they like and use AI to identify what's interesting and apply it to their own writing. - **Broader claim:** Articulating taste is "one of the hardest challenges" in the age of AI; this system helps people "get clearer on what good writing means to them." ### Reasoning about the rebuild - The earlier version (~6–9 months prior) was **programmatic and rigid**: she had to specify every flow and manually tag snippets (voice, structure, substance); insights were rigid. - The new version relies on **compound engineering** and **agent-native architecture**. Her stated reasoning is explicitly secondhand: *"I know compound engineering will make my code better because Kieran wrote it and I know agent native is the future of apps"* — not from her own technical understanding, which she repeatedly qualifies ("technical specifics kind of bounce off my head"). ### Build process (as described) 1. Discussed the idea in chat; had the AI interview her one question at a time. 2. Copied an agent-native architecture guide where Claude Code could find it; instructed "adjust the spec to follow the principles of agent native architecture." 3. Used a compound engineering plugin for code quality/security. 4. Resulting app: clip collection (Leslie Jameson, Amanda Mole, Gia Tolentino, David Foster Wallace), taste analysis, chat-based digging into results, style-guide generation/accumulation, essay feedback against the style guide, and a custom-instructions "TasteMaker voice" in Claude for rewriting drafts. ### Notable feedback example quoted from the tool > "We see solid systems thinking and concrete grounding, but feels more like a straightforward case study than the revelatory excavation your style guide calls for... Missing the revelatory peeling back the structure... lacks philosophical depth and hidden architectures." ### Qualifications and unresolved points (preserved as stated) - She acknowledges the new app is **"a little less pretty"** than the Lovable-built prior version. - Her own writing process is "very chaotic"; it remains open **how much AI output she will keep**, rewrite by hand, or run through other feedback mechanisms — she does not claim the tool replaces her process. - She humorously self-deprecates about buzzword usage ("Do I get points for saying agent native, Dan?") while still asserting the combination "unlocked new capabilities... even as a non-coder." --- ## Closing note from host - Katie will lead an **"AI and writing camp"** soon. --- ## Cross-cutting tensions / open threads - **Secondhand technical confidence:** Katie's enthusiasm rests on trust in others' frameworks (Kieran's plugin, agent-native guide) rather than her own technical comprehension — stated openly, not resolved. - **Tool vs. craft:** Both demos frame AI tooling as enabling, but Katie explicitly leaves open how much AI-assisted output survives her manual editing. # Speaker Attribution & Provenance Record **Video:** What We Built in a Week of Pure Vibe Coding (id: 01a0f15a-aeb3-7a09-9779-596ab64a2592) ## Speakers & Roles | Speaker | Role | Evidence | |---|---|---| | **Unnamed host** | Creator / emcee | Opens "demo day," moderates, introduces each presenter; located in Panama with remote participants. Implied affiliation with **Evergreen** ("how we build internally at Evergreen"). | | **Jack** | Guest presenter; **contributing editor** at "you" (organization name garbled in transcript) | Self-introduction: "I am Jack. I am a contributing editor at you." Previously wrote a piece on AI and creativity. Demoed an agent-activity visualizer. | | **Katie** | Guest presenter; **staff writer**, "extremely AI[...]p editorial team member" | Self-introduction; column "Working Overtime." Demoed **Tastemaker** (formerly Swipe File). Announced as leading an upcoming **AI and writing camp**. | Referenced but non-speaking: **Kieran, Yash, Dan, Naveen** (Naveen noted as skilled at Werewolf; Kieran credited as author of "compound engineering" code; Dan addressed jokingly re: "Andy Dillard"). ## Referenced Sources & Third-Party Claims - **Technologies cited by speakers:** Opus 45, Cloud Code (likely "Claude Code"), Co-work, Claude Code, Lovable, a "compound engineering plugin," and an "agent native architecture" guide. - **Writers named as taste clips in Katie's demo:** Leslie Jameson, Amanda Mole, Gia Tolentino (spelling uncertain), David Foster Wallace, Andy Dillard. - Transcript contains garbled/unclear organization names ("you," "every/Evergreen") — flagged as uncertainty, not corrected. ## Promotional Segments - **Announcement:** Host promotes Katie's upcoming **"AI and writing camp"** — "look out for that." - Katie mentions her **column "Working Overtime"** and Jack references his **prior article on AI and creativity** (self-referential promotion of publication work). ## Availability & Limitations - **Comments:** No metadata supplied; comment availability **unknown**. - **Curated metadata:** None present; only transcript available. - Live stream noted as potentially unstable due to Panama internet; remote participants covering via Slack. - Transcript ends mid-event ("next up" sequence incomplete beyond Katie's demo); further speakers may exist but are not attested in this source. All right, everybody. We're gonna get started. Welcome to demo day. So excited to have you. We are coming in live. Most of us are in Panama in a house together. Um it's really beautiful here, but also the internet is a little bit flaky. So if uh we go out, I'm very sorry. We have a couple people remotely who will be picking up the Slack for us. Um but we're really excited to show you what we did. We have spent the week here on Think Week playing with new technology, thinking new things, hanging out together. Uh we played werewolf last night. Naveen, it turns out, is incredible at Werewolf. Um like just so good. Uh and so yeah, we're just just really excited to be here and show you what we built. I think like we're in this we're in one of those moments where software really changed over the last couple months with Opus 45 and Cloud Code and Co-work. So, we really took the opportunity to see how that might change a lot of how we build internally at Evergreen and I think we have a lot of like really cool demos for you to to show you. Um, so let's get into it. Um, I'm going to have each person uh who demos, they're going to introduce themselves and uh show you what they built. \>\> Um, so yeah, hey everyone. I am Jack. I am a contributing editor at you. You might have read my piece on uh AI and creativity a few weeks ago. Um, so, uh, yeah, for think week, um, the the thing I made was really, you know, inspired by what Kieran and Yash were working on. And it started with the thought, you know, if we're going to be having all these agents working while we sleep, we're going to definitely need different and better ways to actually visualize um, what they're doing while we sleep. So, um, for Think Week, I built \[clears throat\] this. Uh, I'm just going to hit the play button. So what you see here is a uh a scenario that's actually a real research request. Um that's for uh you know an editorial team to identify gaps in our AI coverage. Um so what's happening here is it's using a log file from the actual research request and the events in that log file are acting as triggers to tell the virtual characters on screen where to go. Um so right here, you know, they're coordinating coordinating the research. We got one lead researcher and then four um uh research assistants. Now they're going off and everyone is doing their own thing. Um because a lot of times when we're, you know, when we're actually running the agents, they're doing it in parallel, right? But when we go back and look at transcripts, the transcripts serialize everything that is actually happen in parallel. So this is the way you can actually see um h that happening um in in truly in parallel. Um and so you can do things like uh there's like you know time speed scrub. You can like scrub back and forth. Um you can also like follow a specific agent to kind of hone in on exactly what they're doing. Um and then yeah, let me see what else is there. Yeah. And so at the very end, you know, everyone reconvenes um to discuss their findings. Can speed through this a little bit and then it generates a little research report at the end. That's it. like stunning demo. Amazing work. Thanks for sharing, Jack. \>\> And next up, we've got Katie. \>\> Hi everybody. Uh I am Katie and I'm a staff writer and extremely AIP uh editorial team member here at every you might have heard of my column working overtime. And what I'm going to show you is actually a project that I revived over Think Week um after having previously tried to build it about six to nine months ago. So let me share my screen. Um and what it is, it's called, it used to be called Swipe File and now it's called um Taste Maker. And what it is is it is a tool that um helps you analyze your taste. Um, so as a ghost writer and a sort of editorial support person, a problem I run into a lot with clients that I work with and people I'm supporting is that they know what they like, but they can't explain why. And so the thinking behind this tool is to give people a way to just collect clips of writing that they like and have AI help them say, "Okay, what's interesting about this? Why like how can I apply this to my writing?" So this is the before time version which is very programmat very programmatic. I had to sit and think about every flow that I wanted the app to go through and like specifically request those flows. And so you'll see I've got snippets here and I had to manually tag them with the different things that I liked about them. Voice, structure, substance, all of this I had to think out. Um tags, insights um are also pretty um pretty rigid. And so what I have done now is I'm living in the beautiful world of both compound engineering and agent native architectures. Do I get points for saying agent native, Dan? Um, and I have built a new version of my app. So I'm actually going to go through quickly just the steps that I took to execute this because I think it's interesting for non like I'm deeply non-technical. I technical concepts like the spec like I'm just not a detail person. So like technical specifics kind of bounce off my head and I get confused. But I know compound engineering will make my my code better because Kieran wrote it and I know that agent agent native is the future of apps. So what I have built um so first what I did is I talked to okay being in full screen is actually making my life harder. Let me switch that. Um I started by talking about the idea um in a chat. So this is um here's the initial idea and then I said interview me one question at a time to gather the vision for the project and then somewhere down here you'll see like so I'm just talking about what I wanted to do and then at a certain point you'll see that let me just search for it um adjust the spec to follow the principles of agent native architecture as defined here that's where I took this agent native guide copied for an agent and just dropped it somewhere that my claude code could find. Um, I have the compound engineering plug-in um, plugin set up so that it does the code make look nice and secure and functional. Um, and then this is what I have now. Um, it's a little less pretty because lovable is a little bit better on the design side. Um, but so we've got the clips that I've gathered. Leslie Jameson, Amanda Mole, Gia Tolantino, David Foster, Wallace, you know, the good people. I guess I should get Andy Dillard in here to make Dan happy. And then you can run an analysis. Um, so I just, this is the basic prompt that it will do. Analyze my taste based on my clips. And then down here, you can also chat with this. So, or so you can ask specific like questions that you have. So, this is where the agentic uh functionality comes in because you you're not limited to just, you know, inputting your clips and seeing what the AI puts out. You can dig deeper. You can ask questions. Um and then you can also um you'll see some buttons down here to copy and save somewhere if you want to. You can save clips to your profile. And then you can add it to your either generate a style guide from scratch or add it to your style guide. And this is where you have your style guide laid out. I've actually been working on this a little bit. So, it like it adds and adds and adds as I see new things I want to do. And then if you want to apply this style guide, what you can do, this is I just added today. Um, and so I'm I have this um essay that I'm working on about what board game strat board game mechanics can teach us about the age of AI and um you can get feedback. So, it's going to analyze and here where is, you know, where we all sit here and watch it do its thinking and that's very boring. So I'll just say agent native again um to fill the space. Um but like really seriously like the the the the degree to which the combination of compound engineering and the agent native architecture has just like unlocked new capabilities for me even as a non-coder has just been really exciting because um um it just creates so many opportunities for me to do stuff I could never do before. So, we see solid systems thinking and concrete grounding, but feels more like a straightforward case study than the revelatory excavation your style guide calls for. And then it says, "What's working? What could be stronger? Missing the revelatory peeling back the structure. Um, lacks philosophical depth and hidden architectures." So, these are things that it's noticed about the writing that I gravitate toward that it wants to push me toward adding to um adding to the content. And then the other thing to highlight is you can totally take this um guide and if I go over here to my claude where I still spend a lot of my time in the chat interface um I have set up taste maker voice that has the taste in um this h in the custom instructions and then I could take this same essay and say can you rewrite this in the style defined in the instructions of this project and it will do that. Thank you monologue. Um, and so it's going to look at relevant skills because we've got to make sure that we're getting rid of AI writing patterns and writing according to every style. Um, so that might take a while. Um, but yeah, this is what I've been here every style guide at for grammar mechanics conventions. And then it's just going to start writing um writing the piece. So we'll see, you know, my writing actual writing process for myself is very chaotic. So we'll see how much of this I keep, how much I read by right by hand, how much I run through another version of some other feedback mechanism I've built. But I'm really excited because I think the one of the hardest things challenges that we face in the age of AI is being able to articulate taste. It's something that we've been talking about banging the drum for years at this point and for writing that's so hard to articulate and I think this is just a system that can like help people get clearer on what good writing means to them. So that's Tastemaker. \>\> Amazing Katie and Katie is going to be leading a uh AI and writing camp very soon. So look out for that.