Writings

AI agents are stateless by default. Here's how memory systems actually work, covering the storage patterns, lifecycle triggers, and architecture behind agents that remember you.

How I built a bash-based agent loop that pulls work from Linear, implements features, runs code review, and opens pull requests autonomously.

MCPs give Claude capabilities. Skills teach Claude workflows. Here's the mental model I use to decide which one I need.

LLMs don't have access to the current date, causing issues in time-based analysis. Here's how to fix date and time handling in production LLM systems with explicit context.

How Claude Code's context window works: what consumes tokens (MCP servers, tools, messages), why it matters, and how to manage context effectively.

After 8 months with Claude Code, here's my complete workflow. Learn how I combine Linear, MCP servers, and Obsidian for AI-assisted development that works.

Three design principles for context-efficient MCP servers: filter at source, pre-aggregate data, work creatively. Real reductions: 746k→262 tokens.

Customer obsession can sharpen focus and drive urgency—but left unchecked, it burns out teams and breaks products. Here's how to get it right.

Learn when tech debt is smart, when it’s dangerous, and how to manage it. A clear, startup-tested guide to technical debt from a seasoned engineering leader.

How AI could have helped me generate FDA audit-ready SDS documentation faster while building regulated software.