What We Built in a Week of Pure Vibe Coding

Quick Overview

The team built a new version of their 'Tastemaker' app, which they now call 'Swipe File,' in one week of focused 'vibe coding' by implementing an agent-native architecture that allows AI agents to collaborate on writing tasks, demonstrated by an agent rewriting an essay about board games to adhere to a specific style guide, resulting in a more profound and less superficial output.

Key Points: The team successfully built a new version of their writing assistant, renamed 'Swipe File' (previously 'Tastemaker'), in one week of intensive 'vibe coding.' The core technical shift was adopting an agent-native architecture, allowing multiple AI agents (like Jack, Eleanor, and Rhea) to work in parallel on tasks. A demonstration showed an agent rewriting an essay about board games, applying a complex style guide that emphasized philosophical depth, concrete details, and structural sophistication. The original essay was critiqued for being too much like a case study, lacking the revelatory excavation called for in their style guide. The style guide emphasized principles like maintaining intellectual honesty about complexity, expressing moral urgency without preachiness, and using layered sentence structures. The final revised essay demonstrated the desired style, incorporating deeper analysis and balancing abstract concepts with concrete details.

Context: This video documents a 'vibe coding' week where the creators focused on rapidly developing a new version of their writing assistant tool, formerly known as 'Tastemaker' and now 'Swipe File.' The central concept discussed is moving toward 'agent-native architectures,' which involves structuring the application so multiple specialized AI agents can work collaboratively and concurrently on writing and analysis tasks, rather than relying on a single LLM interaction.

Detailed Analysis

The presentation kicks off with an introduction to the 'Demo Day' and the team's location in Panama, acknowledging potential internet flakiness. The core focus is demonstrating the progress made during a 'vibe coding' week, which centered on building an agent-native architecture for their writing tool, now called 'Swipe File.' Jack explains that this involved using log files from past research requests to trigger virtual agents (Jack, Eleanor, Rhea, Katie) to perform tasks in parallel, such as analyzing traffic data against published content to find coverage gaps, specifically noting under-covered topics like creative AI (Sora, Veo, Kling). The simulation showed agents discussing these gaps, with Eleanor pointing out the under-coverage of high-converting content like 'Thesis' and Jack confirming the data showing a conversion gap. Katie identified topic blind spots in creative AI (video, music, voice) where competitors were active. The demonstration then shifted to Katie showcasing the tool itself, which evolved from a procedural 6-9 month project into a week-long effort due to the agent-native approach. The tool now features a 'Snippets' section for collecting and tagging content, an 'Analyze' section for gaining insights based on these snippets, and a 'Style Guide' editor. Katie demonstrated applying the complex, highly philosophical style guide they developed (which emphasizes deep grounding, structural complexity, and intellectual honesty) to a sample essay about board games. The agent successfully rewrote the essay, transforming it from a straightforward case study into a piece exhibiting the desired philosophical depth and concrete grounding, proving the agent-native system's ability to adhere to complex stylistic instructions.