Builders Unscripted: Ep. 3 - Matias Castello, Product Leader at Alchemy

### TL;DR Matias Castello, Product Lead at Alchemy, shares his journey from a non-engineering background to a tech leader, and how he uses AI—especially Codex and GPT—to radically transform work and personal projects: from code reviews and product docs to having AI autonomously run the entire project management and development pipeline. --- ## Background & Getting Started with AI Matias didn’t come from a formal engineering background, but through curiosity and constant experimentation across multiple domains, he ended up working on consumer products and developer platforms at Facebook, and now leads crypto products at Alchemy. His first real use of AI at work was integrating Codex into Slack to directly modify developer documentation—bypassing the tedious process of running a local site. That early win got the team excited. ## Early AI Applications at Alchemy & The Turning Point The real turning point was a code review mishap. A large migration a few months earlier introduced a race condition bug that caused a minor incident. After fixing it, someone suggested using Codex to re-review that exact code—**and Codex caught the bug**. The team then tested repeatedly, and Codex consistently identified potential issues. Within a week, Matias saw engineers `@mentioning` Codex in pull requests, handling its comments, and iterating. The team overcame the bias that “LLMs aren’t good enough for specialized scenarios.” He also cites DataDog’s data: over one in five incidents could have been prevented by Codex. With model improvements (e.g., GPT‑5.5), he believes that number could exceed half, even nine out of ten. ## Current AI Usage in Daily Workflow Today, Matias uses Codex extensively for product management tasks: - **Writing Product Requirement Documents (PRDs)** - **Analyzing customer feedback** - **Building internal shared skill repositories**—reusable across teams, not only faster and better, but also enabling non‑PMs to accomplish similar tasks. He emphasizes that as an infrastructure company, Alchemy must build tools for two types of “developers”: 1. **Human developers**—100% already using AI‑assisted coding. 2. **Agents**—autonomous AI that require different registration, integration, and execution capabilities. The needs of these two groups are converging, but the differences are still significant. ## The Huge Shift in Entrepreneurship in the AI Era Matias reflects on his own startup experience. Without AI, he manually copy-pasted code to build prototypes and took months with 3–4 engineers to ship an MVP. With today’s AI tools, he believes one person could do the same in a week. He says: **“It has never been easier to build something. This is probably the best time to be a founder—anyone with an idea now has the tools to try.”** ## Personal Project: Letting Codex Work Autonomously Matias built a workflow where he spends very little time in front of the computer while Codex works continuously for hours. He shares a writing app example: ### Writing App Example - He built a Mac app (later ported to iOS) that uses **Codex App Server** and **GPT‑5.5** to help writers polish text. - A global hotkey `Command+Shift+Space` opens a small window for voice input. - Pressing `Command+Return` triggers “Pro Mode,” where AI automatically corrects spelling and elevates tone. - It uses a subscribed ChatGPT account, with inference handled by the App Server. ### Managing Projects with Linear Matias shows a project in Linear (a Mac/iOS app) where **159 completed issues were all created and closed by Codex**—he never manually wrote a single task. The flow: 1. Describe the idea to Codex; Codex creates a plan and breaks it into milestones. 2. Based on an `agent's .md` file that summarizes his preferences (coding style, product goals, etc.). 3. Codex automatically creates tasks, executes them step by step, even tests and validates. 4. He can ask Codex to research the app overnight, build the first ten features (as feature flags), and wake up to toggle them on/off. This approach removes him as the bottleneck for new ideas while keeping control over final quality. ## Tips for Avoiding LLM Assumption Bias Matias believes the biggest gap when using LLMs for coding is that models have to make assumptions, which may differ from the user’s. He solves this by **clarifying all requirements upfront**: writing his preferences and product goals into the `agent's .md` file so that Codex has enough context from the start, reducing unexpected biases. --- **Source:** Builders Unscripted: Ep. 3 - Matias Castello, Product Leader at Alchemy (https://www.youtube.com/watch?v=8QKqENa_eQQ)