3. Building CreatorSignal: From Report to Refinement (LIVE)

This stream picks up where the last one left off. CreatorSignal is an idea validation tool for YouTube creators: you submit a video concept, a research agent hunts across YouTube, Reddit, X, and Hacker News, and you get back a structured report with a score, verdict, and actionable refinement suggestions. The MVP is working. This stream is about making it better.

What happened in this session:

We started by refactoring the research agent. It had grown into a 278-line god object handling the Anthropic client, prompt construction, tool execution, response parsing, and usage tracking all in one place. Claude Code planned and executed the split into five focused collaborators, shrinking the orchestrator to roughly 70 lines.

While that refactor was running, a second Claude Code session acted as a product manager to map out improvements to the onboarding experience. The problem: blank page anxiety. New users land on empty text fields with no idea what to type. We designed a multi-step onboarding wizard.

From there: a live UX prototyping session for AI-suggested video titles. Claude Code generated five iterations of the feature, from a separate card to inline chips to a popover with a sparkle icon and shimmer animation. The popover approach won.

We also discovered that Anthropic’s output_config for structured outputs is cleaner than the tool-use workaround CreatorSignal was using. A GitHub Copilot agent took that on as a parallel workstream via Linear.

The session ended with a three-agent code review on the onboarding wizard. Three separate agents, each with a distinct lens: Ruby/SOLID, JavaScript/Stimulus, and cohesive system review. They found real bugs: strong params silently dropping fields, error recovery gaps, chip state lost on validation failure.