◆ Vibe EngineeringSMIT · AI Agentic Engineering Urdu
Class 11 · 11.3

The pro toolkit: Claude Code, MCP, skills & plugins

Week two turned you from someone who uses an AI IDE into someone who commands a professional coding agent. Let’s reassemble that toolkit so you can see how the pieces snap together — and set up the extensions the finale builds on.

🎯 Goal: re-rack the tools⏱ Read: 15 min🧠 Type: revision

Everything here you have already done once. The goal now is fluency — knowing the tool without looking it up, and knowing why each exists.

01The Claude Code CLI

Claude Code lives in your terminal and treats your whole project as its workspace. The commands you use daily:

CommandWhat it does
/loginAuthenticate your account.
/initScan the project and draft a starter CLAUDE.md.
/contextShow how full the context window is.
/compact · /clearReclaim context, or wipe it for a new task.
/agents · /hooks · /pluginThe pro menus we open up in the finale.
The .claude folder

Almost every pro feature comes back to one folder: .claude/ in your project. Commands, skills, agents, hooks and settings all live there — and because it can be committed to Git, your whole team inherits them. Hold onto this; it is the backbone of Class 12.

02Models & providers

Two separate decisions, easy to conflate: which tool (Claude Code, Cursor, OpenCode…) and which model drives it. And you are never locked in — a provider key repoints the same tool at cheaper or free models.

Frontier

Top quality

Opus-class models for hard reasoning and big refactors. Slower, pricier, worth it when it counts.

Fast frontier

Speed + smarts

Cerebras / fast providers for near-instant responses — great for tight iteration loops.

Free / local

Cheap experiments

OpenRouter, AMP credits, GLM, or Ollama running on your own machine — perfect for practice.

Read the leaderboard

Model rankings shift monthly — check a source like artificialanalysis.ai rather than trusting last quarter’s favourite. The skill is picking deliberately, not memorising a winner.

03Sessions, Git & Ralph loops

Autonomy without a safety net is just risk. Week 2 gave you the net:

Sessions & checkpointsrewind

Move backward and forward through a project’s history; undo a bad turn without losing everything.

Git as the real safety netcommit often

Commit before risky moves. A clean commit is a save point you can always return to.

YOLO mode--dangerously-skip-permissions

Let the agent act without asking — fast, but only over a Git safety net.

Ralph loopsautonomy

Wrap a loop around the agent so it keeps building for hours. Your first taste of hands-off autonomy — and the seed of the hooks and swarms in week 3.

04The big three: MCP, skills & plugins

Class 8’s inflection point. All three extend the agent, and all three build on the one idea underneath everything: tools.

TOOLS — the one idea underneath MCPconnect other people’s toolsJira, GitHub, docs, data Skillsadd expertise as markdownlight, simple, shareable Pluginsbundle it all together+ commands, agents, hooks
A plugin can package skills, commands, sub-agents and hooks — which is exactly what we build by hand in Class 12.6.

05Which one, when?

The rule of thumb

MCP when you need to connect an external system (a Jira board, a GitHub repo, a database). Skill when you want to teach the agent how to do something (call Cerebras, follow your house style). Plugin when you want to bundle and share a whole setup with your team. Reach for the lightest thing that solves the problem — usually a skill.

✓ Key takeaways

  • Claude Code is a terminal agent; the everyday commands and the .claude/ folder are its core.
  • Choose tool and model separately; a provider key repoints you at cheap or free models.
  • Sessions, checkpoints and Git are the safety net that makes YOLO and Ralph loops sane.
  • MCP connects, skills teach, plugins bundle — all built on tools.
  • Reach for the lightest option first. Next: the autonomous workflow and a master cheat-sheet.