./blog
Notes from building and breaking
Write-ups on AI agents, Claude, and the tooling around them. No newsletter, no pop-up. The RSS feed is at the bottom.
RAG that works: chunking, retrieval, and reranking
Most RAG failures are retrieval failures, not model failures. This guide covers the parts that decide quality: how to chunk, how to combine keyword and vector search, when to rerank, and how to measure retrieval separately.
Long context or RAG? How to choose
Bigger context windows didn't kill retrieval, and retrieval didn't make long context obsolete. This post gives you a practical way to pick between them, or combine them, based on your data and your questions.
Prompt caching: cheaper and faster LLM calls
Agents resend the same long prefix on every step. Prompt caching lets the provider reuse that work at a fraction of the price and with lower latency. Here's how it works, how to set it up, and what quietly breaks it.
Prompt engineering that still works with modern models
Modern models need less trickery and more clarity. These are the prompting habits that keep paying off: say what you want, give the reason, show examples, structure the input, and test changes.
Context engineering: what to put in the window and what to leave out
Prompt wording matters less than what else sits in the context window. Context engineering is the practice of choosing, trimming, and refreshing that content on every model call, and it's where agent quality is won.
Build a small MCP server in Python, start to finish
A hands-on walkthrough of writing a Model Context Protocol server with the Python SDK: tools, a resource, a prompt, testing with the Inspector, and connecting it to Claude Code or a desktop app.
Model Context Protocol (MCP) explained for developers
MCP is an open standard that lets AI apps discover and call tools, read resources, and use prompts from any server that speaks it. Here's how it works, when it's worth using, and where the risks are.
How to design tools an LLM can actually use
The quality of an agent's tools sets a ceiling on the agent. Here are the design rules that matter most: fewer tools, clear names, honest descriptions, compact outputs, and errors a model can act on.
Build your first agent with the Claude API in Python
A complete, runnable agent in about fifty lines of Python: two read-only tools, a loop, a step limit, and logging. No framework required, and every line explained.
Prefer a feed reader? New posts show up in the RSS feed as soon as they're published.
RSS feed