Descriptions:
VelvetShark introduces “Compound Engineering,” a structured five-stage AI coding workflow designed to improve with each session rather than reset to a blank slate. The core distinction from ordinary AI-assisted coding is a final compounding step where findings from code review—style preferences, architectural decisions, recurring bugs—are written back into project configuration files and applied automatically to every future task, creating genuine cumulative improvement over time.
The five stages are: brainstorm (Claude interviews you to surface edge cases and produce a spec), plan (agents research the codebase and pull live docs via the Context7 MCP server, which covers over 100 libraries), work (agents execute the plan autonomously), review (specialized parallel agents critique the code from perspectives including code quality, simplicity, and domain-specific standards like DHH-style Rails philosophy), and compound (session insights are persisted to config). The video demonstrates the full cycle by adding AI-generated header images to a real Next.js/MDX blog using the Nano Banana Pro image generation API.
The workflow plugin installs into Claude Code with two commands, providing 27 agents and 20 slash commands. Three parallel review agents—including one named after Kieran, a workflow co-creator—identified blocking issues and recommended generating one image instead of two before any code was written. One practical catch surfaces: Claude assumed `.env.local` for environment variables because the actual `.env` file was gitignored, illustrating the value of the review stage. The creator’s key principle: spend 80% of effort on planning and review; coding becomes nearly automatic when the plan is solid.
📺 Source: VelvetShark · Published January 23, 2026
🏷️ Format: Workflow Case Study
