RecapFlow : Avril 7th Coaching call analysis
π SUMMARY This week's call covered demos, workflow strategies, and tool discoveries. Eric Li demoed Kira, a natural language Text-to-SQL platform with a standout Slack integration. Patrick shared updates to his Claude Code plugin ecosystem and an Obsidian-based knowledge management setup. Ryan walked through a rebuilt SEO-optimized agency site and a hands-free lead capture app. The group also discussed enterprise agentic development workflows, the Claude Code source leak and its Rust rewrite, Gemma 4 for local/private AI, and the Claude plus Codex hybrid development pattern. π‘ KEY INSIGHTS Claude + Codex hybrid workflow is delivering extraordinary value Patrick runs Claude Max ($100/month) alongside the OpenAI Codex plugin ($20/month) and estimates he is getting roughly $5,000/month in API value. The plugin adds slash commands into Claude Code sessions including /codex review, /codex status, /codex rescue, and an adversarial review function that actively tries to break your application and reports vulnerabilities. You can configure a completion hook so Codex automatically reviews every Claude output in a loop until no issues remain. Obsidian plus Claude Code as a lightweight RAG Using Andrej Karpathy's prompt (linked below), you can turn an Obsidian vault into a self-organizing knowledge base managed by Claude Code via terminal. It auto-classifies and backlinks documents and performs comparably to a light RAG system for up to a few hundred documents with no additional infrastructure. Enterprise SDLC adapted for agentic development Patrick's team records design meetings, transcribes them, and has Claude Code convert them into an intent document split into independent workstreams for parallel development across a team of four. Sprint planning transcripts become the workstream specs β a scalable model for teams moving beyond solo agentic development. User-driven development loop (Ty's model) Ty built a feedback hub where end users describe issues with screenshots or video. Claude Code asks clarifying questions until it reaches 85% confidence, then autonomously builds a fix in a headless session, creates a PR, and Ty only approves and merges. The Codex plugin validates before the PR is created.