A Claude Code session is your working context — conversation history, file state, permissions, and memory. Understanding how sessions work and how to manage them unlocks the ability to parallelize work, switch contexts without losing state, and recover from mistakes instantly. This guide covers session fundamentals, parallel work patterns, and five developer habits that compound over time. What Is a Session? A session is: Directory-scoped, not branch-scoped — tied to a folder location, not a git branch Persistent across context limits — survives /compact and /clear Shareable — /rename and /resume to switch between named sessions Isolated — each session has its own context window, permissions, memory Critical distinction: Sessions follow your directory, not your branch.

A sub-agent is a specialized Claude instance that runs in an isolated context window with its own allowed tools, memory, and instructions. While your main Claude session handles the conversation, a sub-agent spins up to handle a specific task — investigation, code review, security audit, testing — in parallel with your work. Sub-agents excel at tasks that require: Reading many files without cluttering main context Specialized expertise in one domain Fresh perspective (the agent isn’t biased by earlier conversation) Parallel execution (multiple agents working simultaneously) Sub-Agents vs Skills vs Hooks Before diving into sub-agents, understand the distinctions:

A Skill in Claude Code is a specialized workflow that Claude can invoke automatically or you can trigger manually with a slash command. Unlike CLAUDE.md rules (which advise on every decision) or Hooks (which enforce deterministic actions), Skills are progressive disclosure — deep knowledge that loads only when relevant, keeping your base context lean while providing specialized expertise when needed. This guide covers how Skills work, when to use them, and how to build real Skills that serve developers, testers, and DevOps engineers.

Skills, agents, and hooks you add to .claude/ are powerful — but they are locked to one project. Every time you start a new repo you copy the same files, maintain them in multiple places, and drift out of sync. Claude Code plugins solve this: a plugin is a shareable, versioned package that carries all your customisations and can be installed in any project with one command. This post covers what plugins are, when to use them, and how to build a real one from scratch — a DevOps helper that ships a deployment skill, a pre-deploy safety hook, and an MCP server connection to your Kubernetes cluster.

By default, Claude generates its response token by token without any deliberate planning step. For most tasks — answering a question, writing a function, explaining a concept — this is fine. The response comes quickly and it is good. For some tasks, it is not enough. Complex multi-step reasoning problems, ambiguous architecture decisions, intricate security analyses — these benefit from Claude thinking through the problem before committing to an answer. That is what extended thinking and adaptive thinking provide.

Building an AI agent that runs autonomously — browses the web, executes code, reads and writes files, persists memory across sessions — requires infrastructure. You need a sandbox, a process that can run for hours without your web server timing out, and a way to resume from where you left off after a network hiccup. Claude Managed Agents, launched in public beta in April 2026, offloads all of that to Anthropic.

Picking the wrong Claude model is expensive. Opus on every task costs 5x more than Sonnet for comparable results on most work. Haiku on a complex reasoning task produces worse output than just asking Sonnet. And if you are still using models from early 2025, some of them are deprecated — or will be soon. This guide covers every current Claude model, what each is good at, how much they cost, and a concrete decision framework for choosing the right one.

If you are calling the Claude API repeatedly with a large system prompt, a big document, or a long codebase context — and you are not using prompt caching — you are paying full price every time for content that has not changed. Prompt caching stores a prefix of your prompt server-side and charges 90% less to read it back on every subsequent request. For applications that repeatedly process the same context, this is the single highest-impact API optimisation available.

Spring AI 1.0 GA shipped in May 2025. It brings the Spring programming model to AI development: a unified ChatClient API that works across Claude, OpenAI, Gemini, Ollama, and Azure OpenAI — switching AI providers is changing one dependency. This guide builds a complete RAG (Retrieval-Augmented Generation) application that answers questions about your documentation using any AI provider. What Is RAG? A large language model (LLM) knows everything in its training data but nothing about your specific documents, code, or business data.

Most AI assistants are chatbots. You ask, they answer, the interaction ends. The interesting shift happening right now is treating AI as an autonomous worker — something that runs on a schedule, produces real artifacts, and delivers results without you being in the loop. This post walks through building that kind of assistant: a background agent that runs weekly, researches a set of topics relevant to your work, and delivers a structured briefing via Telegram or email.