I design the systems around Claude Code: project structures, automation architectures, and verification pipelines that make AI-paired development work at production scale. That means CLAUDE.md project files over 900 lines that let any session pick up cold without losing context, and intent-based skill architectures where markdown files describe goals and constraints instead of step-by-step scripts, letting the AI figure out implementation.

My flagship system runs an Etsy shop end-to-end with 34 specialized skills and three tiers of learned memory. The intent-based architecture cut automation code 74% while increasing capability. I've done a full architectural rewrite in 6 days when the original wasn't scaling, and scrapped 2 weeks of Chrome extension work after 8 PRs when a simpler approach was clearly better. 16 months and 260+ sessions of building, breaking, and rebuilding. I also maintain zen-mcp-server (open source, 50+ models, 765 commits).

Every blog post I publish goes through a 4-agent review pipeline: fact-checker, technical reviewer, editorial reviewer, and visual QA. I built it because the default LLM workflow is ship-first-verify-never, and that's not an option when people act on what you publish. Business metrics get cross-checked against raw data sources. Claims get traced to primary sources. I've caught errors in my own drafts, retracted incorrect findings after publishing, and corrected community misinformation that was leading people to wrong fixes .

Concrete example: my verified reference for multi-agent Claude Code development cites 50 claims against primary sources. I caught 4 errors in my own draft during the verification pass. I've also written about why agents can't evaluate their own work , backed by NeurIPS and ICLR research. If the output matters, I'd rather find my own mistakes than let a customer find them.

When something breaks and there's no documentation, I trace the root cause through log files, undocumented APIs, and system internals, then publish the fix. I've diagnosed Anthropic's Cowork startup freeze , its file upload failure , and 6 distinct VM failure modes hiding behind one generic error. I've filed issues upstream on Google's Gemini SDK (silent parameter stripping, no error message) and Anthropic's GitHub issue tracker.

I reverse-engineered Cowork's undocumented session format from 146 files and mapped where it actually stores conversations across 195 sessions, surfacing per-session cost data that no community tool captures. If standard troubleshooting fails, I keep going until I find the actual cause.