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

Key Concepts

• AI-Driven Development: Using Large Language Models (LLMs) to automate coding, project management, and research.
• Codex App Server: A backend infrastructure that allows developers to integrate LLM capabilities (like GPT-5.5) into custom applications and workflows.
• Autonomous Agents: Software entities capable of performing tasks, making implementation decisions, and interacting with infrastructure independently.
• Skills/Agentic Workflows: Modular, reusable sets of instructions or scripts that allow AI to perform complex, multi-step tasks (e.g., research, planning, coding, testing).
• Feature Flagging: A technique used to toggle experimental features on or off within an application, allowing for rapid testing and iteration.
• Computer Use: An AI capability that allows models to interact with a computer interface (mouse, keyboard, browser) to perform tasks as a human would.

---

1. Evolution of AI in Professional Workflows

Matias, a product leader at Alchemy, highlights how AI has transformed the software development lifecycle.

• Initial Adoption: The first major use case was automating documentation edits via Slack, followed by automated code review.
• The Turning Point: The team realized that AI could retroactively identify bugs and race conditions in complex codebases. This shifted the team's perception of LLMs from "experimental" to "professional-grade tools."
• Agentic Shift: Developers are now viewed as "agents" consuming infrastructure. The platform must now support both human developers and autonomous agents that can sign up, integrate, and execute tasks on the blockchain without human intervention.

2. Personal Productivity and "Building While Sleeping"

Matias describes a methodology for building software without being a constant bottleneck:

• The "Skills" Framework: He maintains a shared repository of "skills"—pre-defined instructions that allow the AI to perform specific PM tasks (writing PRDs, analyzing feedback) or coding tasks.
• Asynchronous Development: By using an agents.md file that defines his personal coding preferences and project goals, he can prompt the AI to build entire features or research competitors while he is away from his computer.
• Experimental Loops: He instructs the AI to research competitors, generate a list of features, and implement them behind feature flags. This allows him to wake up to a list of toggles, enabling him to test and approve features without having written the code himself.

3. Technical Implementations and Real-World Applications

• Writing Assistant (Mac/iOS): A custom app built using the Codex App Server that uses voice dictation to rewrite text into professional formats.
• Discord/Open Claw Integration: A bot named "Lou" that allows Matias to trigger coding tasks via Discord from his phone while on the go.
• Apple Watch Integration: A voice-memo-to-code workflow where a short recording triggers a transcription, intent analysis, and a GitHub pull request, demonstrating the potential for "on-the-go" development.
• Computer Use: Matias demonstrated using AI to SSH into a Raspberry Pi and automate tedious data entry tasks, noting that the model successfully navigated the UI and performed the actions autonomously.

4. The "Snapchat for Cats" Eval

Matias uses a 10-year-old hackathon project—an app that takes photos of cats when they tap a laser dot—as a personal benchmark for model capability.

• Evolution of Capability: What once took a team of five people 24 hours to build can now be "one-shot" (built in a single prompt) by current models.
• UI Generation: He demonstrated a new workflow where he generates UI designs via image prompts, then instructs the AI to implement those designs into the codebase, effectively separating design ideation from technical implementation.

5. Key Arguments and Philosophy

• The "Assume It’s Your Fault" Principle: Matias argues that when an LLM fails to produce the desired result, the user should assume the communication was unclear rather than blaming the tool. This requires putting one's ego aside to refine prompts and context.
• Democratization of Building: He emphasizes that the barrier to entry for founders has collapsed. A product that once required 15 engineers and 18 months can now be prototyped by a single person in less than a week.
• The "Implementation Detail" Vision: The ultimate goal is to reach a point where the actual coding process becomes an "implementation detail," where the user simply defines the intent and the AI handles the execution.

Conclusion

The conversation underscores a fundamental shift in software development: the transition from manual coding to orchestrating AI agents. By building modular "skills," utilizing infrastructure like the Codex App Server, and adopting an experimental mindset, builders can significantly increase their output. The core takeaway is that the primary constraint for modern builders is no longer technical skill, but the ability to clearly define intent and manage the AI agents executing the work.