A routine Claude Code PR review burned 48% of my 5-hour quota. Sub-agents and conversation length were red herrings. The real cause was a 51.9KB MEMORY.md being re-injected into the system prompt every turn. The cost equation, what got fat, and the index pattern that fixed it.
Cloudflare's isitagentready.com scores sites across four categories: Discoverability, Content, Bot Access Control, Capabilities. A field report of what iammattl ships in each, what's missing, and one thing the rubric doesn't check that's still worth doing.
Two parallel implementations of the same idea: one small improvement to my site, every hour, forever. A Claude Code scheduled task running locally and a five-agent pipeline on Cloudflare Workers called Vigil. Same rule at the top of both: ONE improvement per run, never batch.
A day at Cloudflare Connect London during Agents Week. Sandboxes, MCP Server Portals, Cloudflare Mesh, Moltworker, Kilo Code, and AI Crawl Control. I deployed an AI agent and hardened an MCP server before the day was over.
Skills tell Claude how. MCP servers give Claude access. Agents let Claude work autonomously. The final post in the Extending Claude Code series covers subagents, agent libraries, multi-agent orchestration, and the trust question.
Skills tell Claude how to do things. MCP servers tell it where to look. How I use Cloudflare, GitHub, Context7, Chrome DevTools, and a custom TrueNAS server to give Claude direct access to real infrastructure.
Skills and plugins let you package expertise so Claude works consistently across projects. Here's how I built an SEO suite with 13 skills and 6 subagents, and the patterns that apply to any workflow.
Most people use Claude like a search engine with better grammar. Here's how to actually make it useful.
After building multiple projects with Claude Code — from Terraform modules managing 400+ Cloudflare zones to a full-stack price comparison site — here's what actually works when using AI as a pair programmer.