I write code with Claude Code and Opus, and content with Codex. I rarely reach for third-party skills — CodeGraph finds the relevant code files, and the rest I build myself:
a /kanban skill I built for project tracking. The agent proposes new tasks and solutions on its own, and it learns from my feedback so it gets better over time. It splits large tasks, prunes its memory regularly to keep the context clean, and pushes me on requirements I've left out. No RAG. It covers the three sides of my day: building, growth and marketing, and validation — the social posts I use to gather feedback.
Content writing
a workflow and a set of verifiers for things like SEO blogs, Reddit posts, and usage guides. Each verifier is a single-pass subagent that catches flaws in the writing. It often over-corrects, so I'm still refining the loop.
Skills for skills
when I close a session, I run a command that turns my feedback into requirements for the verifier skills, so the loop gets smarter over time — the same idea as the Hermes agent.