Blog
Thoughts on product engineering, MCP servers, AI tools, and developer experience.
One MCP Tool Call Can Burn Your Whole Token Budget
MCP tool responses flooding the context window? How the Opik MCP server cuts token usage: three detail levels, a token count per answer, no silent cuts.
MCP Tool Design for AI Agents, Not API Endpoints
Long MCP tool lists waste LLM context and confuse AI agents. An MCP server design that swaps endpoint-shaped tools for one execute and one describe tool.
Stop Using Claude Code on Defaults
The Claude Code settings.json changes I use to cut token costs, stop hallucinations and skip permission prompts: model routing, LSP, Gemini delegation and more.
Agentic UX Primitives: The Frontend Patterns Nobody Taught You
Agentic UX patterns for AI agent UIs: streaming with intent, human-in-the-loop gates, reasoning visualization and confidence indicators.
Context Engineering Ate Prompt Engineering
Context engineering vs prompt engineering: why the context you give AI coding tools (CLAUDE.md, MCP servers, skills) matters more than the prompt.