AI in practice
Current tools and trends, and how to use them well. Written for working developers, with the JSON side of each topic covered because that is where things break.
8 articles, newest first
-
How to Make Your AI Use Fewer Tokens, and What It Adds Up To Over a Year
Most AI spend is not the answer, it is everything sent around the answer. Six instruction and habit changes that cut token use, why each one works, and a worked estimate of what they add up to across a year of daily use. No magic, just less waste.
-
CLAUDE.md, AGENTS.md and Custom Instructions: Teaching Your AI Who You Are
Every serious AI tool now reads a Markdown file or a settings box before it does anything. Done well, that file turns a generic assistant into one that knows your stack, your rules and your taste. A guide to the instruction files in Claude Code, Codex, Cursor and Copilot, the custom instruction settings in ChatGPT, Claude and Gemini, and what actually belongs in them.
-
Tokens: The Strange Currency Every AI Model Runs On
A model does not see words, letters or sentences. It sees tokens, and tokens are the unit it thinks in, pays in and runs out of. A tour of the odd consequences: why emoji are expensive, why Chinese costs more than English, why a model can write a poem but cannot count its letters, and where your money actually goes.
-
AI Agents in Everyday Life: What They Do Well and What Not to Trust Them With
Agent mode in ChatGPT, Gemini and Claude will browse, book, fill forms and act on your behalf. A practical look at what these agents are good at in 2026, the tasks where they quietly fail, how to give them access safely, and how to think about a tool that acts instead of answers.
-
How to Get Reliable JSON Out of an LLM
Language models are text generators, and text is not always valid JSON. The techniques that actually work in 2026: structured output modes, JSON Schema, prompt patterns, validation and repair, and the failure cases you still need to handle.
-
Run an LLM on Your Own Laptop with Ollama: A Practical 2026 Guide
Open weight models are good enough now that a laptop can run a capable assistant with no subscription and no data leaving the machine. Which hardware you need, which models to pull, how to use the local API from code, and where local models still fall short.
-
What Is MCP? The Model Context Protocol Explained for Developers
MCP went from an Anthropic side project to a standard supported by OpenAI, Google and Microsoft in about a year. What problem it solves, how servers, clients, tools and resources fit together, a minimal server, and the security issues you need to know about.
-
AI Coding Agents in 2026: What Actually Changed for Developers
Claude Code, Codex and Cursor have converged on the same idea: an agent that reads your repo, runs commands and edits files on its own. What that looks like day to day, where it saves time, where it wastes it, and how to work with one without losing control of your codebase.