Home›Practice
Practice
Practice
Hands-on guides and first-hand notes on agent frameworks, skills and AI tools.
Total 217
Using Claude Code in VS CodeInstall the Claude Code extension in VS Code, sign in, review proposed edits, resume sessions and fix common setup problems, without leaving the editor.Connecting MCP servers to Claude CodeAdd an MCP server to Claude Code with claude mcp add, check its status, choose a scope, approve project servers and fix the usual connection errors.Claude Code headless mode with claude -pRun Claude Code without its interface using claude -p: exit codes, piped input, JSON output, permission flags, bare mode and session resumption for scripts and CI.Claude Code subagentsSubagents are assistants that handle one kind of task in their own context window. This guide shows how to write a subagent file, where to store it, how to limit its tools and permissions, and three ways to invoke it.Claude Code hooks in practiceHooks are shell commands that Claude Code runs at fixed points in its lifecycle. This guide walks through two working examples, auto-formatting after edits and blocking edits to protected files, and explains how to debug a hook that does not fire.Claude Code best practicesDefaults handle most work. When a project has rules the model cannot see, add them to configuration: tell Claude what it lacks, give it a check it can run, and enforce non-negotiable rules with hooks and permissions.Common Claude Code workflowsEveryday Claude Code recipes for reading code, fixing bugs, writing tests and opening pull requests, plus worktrees, plan mode, subagents and scripted use for keeping sessions focused.Fixing Claude Code installation problemsClaude Code install failures usually come down to PATH, network or proxy problems, certificates, leftover older installs, or a broken sign-in. This guide maps common error messages to the fix for each.Semantic search with Postgres and pgvectorBuild semantic search on Postgres with pgvector: install the extension, store embeddings as halfvec, create an HNSW index, and re-embed only the text that changed. Shows the setup behind TellWorld's question-answering feature.Batch AI illustrationMake a series of AI cover illustrations with one shared style, convert them to small WebP files, and check each one for stray text with Apple's Vision framework. Covers the prompt file, cwebp options and upload order.A fact-checked AI writing pipelineHow to build an AI writing process that does not invent facts: separate gathering from writing, attach a source to every claim, ask the model to cite, and run checks before publishing. Covers citations, retrieval limits and evaluation.Keeping agent costs downThree levers for cutting the cost of AI agents: choose the cheapest model that fits each task, cache repeated prompt context, and batch work that is not urgent. Read current Claude and GPT prices and set up a first cache breakpoint.Splitting work across agentsCompare three ways to split work across AI agents: Claude Code subagents, OpenAI Agents SDK handoffs and LangGraph orchestrator-worker graphs. Set up a first subagent file, and see the context and cost limits to plan for.Project instruction files for agentsCoding agents read project instruction files before they work. This guide explains how CLAUDE.md, AGENTS.md, Cursor rules and GEMINI.md differ, how to write one an agent can follow, and how TellWorld uses its own CLAUDE.md.Browser agentsBrowser agents let a model click, type and read web pages. This guide compares three ways to do it, Claude in Chrome, Playwright MCP and the computer use tool, with setup steps and the safety limits of each.Getting started with Codex CLIOpenAI's Codex CLI runs an AI agent in your terminal. This guide covers installing and signing in, interactive and codex exec use, sandbox and approval settings, and how we use it to generate cover images.LangGraph quickstart: build a tool-calling agentLangGraph builds agents as explicit graphs of state, nodes and edges. This quickstart wires a small model-and-tool loop with Claude in one Python file, then shows how to add memory and pause for human input.Building a RAG knowledge baseRAG (retrieval-augmented generation) lets a model answer from your own documents. This guide sets up a knowledge base in Dify without code, then builds the same idea in a short LangChain script, and covers chunking, retrieval and limits.Running open models locally with OllamaOllama runs open models such as Qwen3 and DeepSeek-R1 on your own computer and serves them through a local API. This guide covers installing it, running a first model, calling the OpenAI-compatible endpoint, and the memory limits you will hit.Agent tools available in ChinaFive agent tools with Chinese interfaces or domestic sites: Dify, Coze, WorkBuddy, TRAE and Tongyi Lingma. How to start each one, how their plans and credits work, which models they offer, and where each fits.Coding agents comparedClaude Code, Codex CLI, Gemini CLI and TRAE compared side by side: installation, project instruction files, models, plans and prices, and which one fits which situation.Building an MCP server from scratchHow to build a small MCP server in Python or TypeScript: the three building blocks, the two transports, testing with the MCP Inspector, and connecting the server to Claude Desktop and Claude Code.How to write an Agent SkillHow to write an Agent Skill: the SKILL.md format, the name and description rules, optional folders, where skills live, how agents load them, and how TellWorld's news update skill is built and run.Getting started with Claude CodeA hands-on guide to Claude Code: installing it, starting a session, writing CLAUDE.md, adding a subagent, reading permission modes and hooks, and how our own setup uses these pieces in practice.