Beyond the Prompt: Claude Code

Takeaways

Core Boris principle: give Claude a verification loop (tests, linters, runnable checks)—claimed 2–3× quality vs. human-only feedback.
Workflow: explore → plan (Plan mode) → code; use a second session to review plans as staff engineer without implementation bias.
Reference with @files and pipe logs—exact context beats vague descriptions; delegate like an engineer, not pair-program line by line.
On mistakes: “Update CLAUDE.md so you don’t repeat this”—Claude is strong at self-authored rules that compound over time.
.claude/ holds skills, subagents, plugins; underused commands (/goal, /insights) and MCP wiring separate casual from power users.

High-signal playbook for tuning this vault’s Claude Code + skills setup—especially verification hooks and CLAUDE.md feedback loops.

Which verification patterns (tests vs. skill dry-runs) fit markdown vault changes best?