Prompt Engineering, Skills & Agents

The Basics: What Makes a Good Prompt?

Before and After

Core Prompting Techniques

Chain-of-Thought (CoT)

Instead of asking for an answer directly, ask the model to think step by step. This dramatically improves accuracy on reasoning tasks because it forces the model to “show its work” and catch mistakes mid-generation.

// Without CoT
"What is 17 × 23?"

// With CoT — much more reliable
"Calculate 17 × 23. Think step by step before giving the final answer."

Few-Shot Prompting

Give the model examples of what you want before asking it to do the task. This is called “few-shot” learning. The model infers the pattern from the examples.

Input: "The weather today is great."
Sentiment: Positive

Input: "I missed the bus and now I'm late."
Sentiment: Negative

Input: "The package arrived on time."
Sentiment: ???

System Prompts

Most LLM APIs accept a “system prompt”, a special instruction that sets the model's behavior before the conversation starts. Think of it as the model's job description and personality.

You are a senior TypeScript engineer at a startup. You write clean, idiomatic code.
You always:
- Use TypeScript strict mode
- Prefer functional patterns
- Add brief comments on non-obvious logic
- Flag potential security issues

You never add unnecessary boilerplate or over-engineer simple solutions.

Skills (Tools)

A plain LLM can only generate text. But modern AI systems can be given tools, also called skills, that let them take real actions in the world.

When you use Claude Code, you're using an LLM with tools. Claude can read your files, run terminal commands, edit code, and remember context, all because it has been given those capabilities as tools.

MCP: Model Context Protocol

Tools give AI agents abilities. MCP (Model Context Protocol) is the standard that lets those tools be shared, discovered, and connected across different AI systems. Think of it as a universal plug: instead of every AI tool being custom-wired to one specific AI, MCP lets any compatible AI agent connect to any compatible tool server.

To enable MCP servers in Claude Code, add them to your settings.json file (found at ~/.claude/settings.json). The Claude Code docs maintain a growing list of community MCP servers you can add in minutes.

Claude Code Slash Commands & the Skill Builder

Beyond tools Claude uses automatically, Claude Code ships with built-in slash commands, pre-built workflows you trigger yourself by typing / in the terminal. Think of them as keyboard shortcuts for common dev tasks.

The Skill Builder: Create Your Own Slash Commands

Claude Code lets you create custom skills, your own slash commands that encode repeatable workflows. A skill is just a Markdown file you save to ~/.claude/skills/. Once saved, it shows up in the / menu just like a native command.

This is a huge win for beginners: instead of re-explaining your preferences every session, you bake them into a skill once and reuse them forever.

---
name: explain-this
description: Explain the current file to me in plain English
---

Read the file I'm currently working on and explain it like I've never seen it before.
Walk through:
1. What this file does
2. How it's structured
3. The key functions or components and what they do
4. What I'd need to understand to safely modify it

Use plain English. No jargon without definition.

Save that file, then type /explain-this in any Claude Code session and it just works.

Agents

An AI agent is an LLM that can act autonomously in a loop: observe → think → act → observe → think → act, until it completes a goal.

Claude Code is an agent. When you tell it “build me a login page,” it reads your codebase, plans the component, writes files, runs the linter, fixes errors, and reports back. That's the agent loop running.

Multi-Agent Systems

You can chain agents together: one agent plans, another writes code, a third runs tests, a fourth handles deployment. Claude Code's “subagent” system works exactly this way, spinning up specialized agents for different parts of a task.

Claude Code: An Agent Built Into Your Terminal

Claude Code is a CLI tool that turns your terminal into an AI-powered development environment. It's not just autocomplete. It's a full agent with access to:

• Your entire codebase (reads files, follows imports)
• Your terminal (runs commands, sees output)
• Git (reads history, creates commits)
• A persistent memory system (remembers things across sessions)
• Sub-agents it can delegate tasks to

It uses skills (configured in settings.json) to control what tools it's allowed to use automatically vs. what requires your approval.

Claude Code vs. Cursor vs. GitHub Copilot

Every beginner asks this. They're all AI coding tools, but they're built differently and shine in different situations.

They're not mutually exclusive. Many developers use Cursor for in-editor work and Claude Code for bigger agentic tasks. For this guide we're focusing on Claude Code because it teaches the agent model most clearly.

Prompt Templates for Developers

Steal these. Adapt them. Use them every day.

Review this code for bugs, security issues, and performance problems. Be specific: point to line numbers and explain why each issue matters. Then suggest improvements.

[paste your code here]

Explain this code to me like I've never seen it before. Walk through what it does, why it's structured this way, and what I'd need to know to modify it safely.

[paste your code here]

I'm getting this error: [paste error]. Here's the relevant code: [paste code]. Here's what I've already tried: [what you tried]. What's causing this and how do I fix it?

I want to build [feature]. My stack is [stack]. Before writing any code, give me a step-by-step implementation plan. Point out the hard parts and any libraries that would help. Only then write the code.