Published: 2026-04-13

The Ultimate Claude Code Guide: Skills, Sub-Agents, and MCP Servers

Most people open Claude Code, start chatting immediately, and only reach about 20% of its actual capability. Tech With Tim walks through his real professional setup — covering custom skills, parallel sub-agents, and MCP server connections — showing how Claude Code becomes extremely powerful once customized to your workflow.

Source video

"The Ultimate Claude Code Guide | MCP, Skills & More" by Tech With Tim — Watch on YouTube →

Key Takeaways

Default chat mode captures only ~20% of Claude Code's capability — skills and sub-agents are the unlock that separates power users from casual users.
Custom skills are reusable instruction modules: create once, invoke with a slash command, and Claude executes a consistent multi-step workflow every time. They're the closest thing to "programming Claude Code" in plain language.
Sub-agents run in parallel — assign different tasks to specialized agent instances and collect results. Essential for large codebases where different components need different expertise.
MCP (Model Context Protocol) server connections let Claude Code interact with external tools — databases, APIs, browsers — natively inside the agent loop without leaving the session.
Tim's production setup: a CLAUDE.md with project context, 3–5 domain-specific skills, 1–2 MCP servers, and sub-agents for testing and reviewing. This setup took one afternoon to build and pays off on every project after.

The Professional Setup in Practice

Tim walks through his actual workspace configuration. The CLAUDE.md file establishes project context so Claude doesn't re-ask basic questions. Skills handle recurring workflows: a /test skill that runs tests and summarizes failures, a /review skill that checks a PR against internal standards, a /deploy-check skill that runs pre-deployment validation.

For MCP servers, Tim uses a GitHub MCP server (read PR and issue context directly) and a database MCP server (query production read replicas for debugging). Both are configured once in the Claude Code settings and available in every session.

Sub-agents shine on large refactors: spawn one agent to handle the API layer, another for the UI, another for tests — then collect and reconcile their outputs. The coordination cost is worth it above ~500 lines of change.

Related on OpenClawDatabase

Claude Cowork Skills Guide — how skills work across Claude platforms
Skills Database — community-contributed skills for common workflows

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Commands & Code Shown

/model

Purpose: Switch the model for the session (Sonnet default, Opus for hard work, Haiku for simple or bulk work)

When to use: Especially on the $20 plan, where defaulting to Opus burns usage fast

/insights

Purpose: Generates a local HTML report on how you use Claude Code: what works, what doesn't, mistakes, charts

When to use: Once in a while, to audit your own habits

/cost

Purpose: Shows session spend when running on the API

When to use: API billing only; shows nothing useful on a subscription

/context

Purpose: Breaks down what is filling the context window (system prompt, tools, skills, MCP tools, messages)

When to use: In long sessions, and to find MCP tools you are loading but not using

/compact

Purpose: Summarises the conversation so far and keeps going with a smaller context

When to use: Near the context limit, when you want to keep the same session

claude mcp add-json github '<server JSON with your GitHub personal access token>' --scope user

Purpose: Adds the GitHub MCP server for every project (user scope) instead of only the current one

When to use: Run in a normal terminal, not inside a Claude Code session

/mcp

Purpose: Lists MCP servers and their status; reconnect or re-authenticate from here

When to use: After adding a server, restart Claude Code and check it shows as connected

/skills

Purpose: Lists available skills

When to use: To confirm a newly created skill was picked up

Reproducible steps

Create a GitHub personal access token
GitHub → Settings → Developer settings → Personal access tokens → Fine-grained token → create, giving it repo access. Paste it into the MCP add command.
Know the three MCP scopes
local (default: only you, only this project), project (shared via .mcp.json), user (every project). Project-specific config lives in the .claude folder inside the repo.
Install plugins from the Claude desktop app
Code tab → Customize → Personal plugins → Browse plugins. Context7 (free, live documentation), Playwright (browser automation so Claude can test sites it builds), Superpowers (pre-built agents and skills, heavy on context).
Cycle permission modes with Shift+Tab
Switches between Accept edits, Plan mode and the default ask mode.
Create a skill
Ask Claude to create it, e.g. a code-review skill with a fixed output format (security, error handling, performance, code quality, testing; severity and location per finding). It is written to .claude/skills/code-review/SKILL.md. Restart the session, then invoke with /code-review, or just ask for a code review and Claude picks the skill automatically.
Move a skill to global scope
Ask Claude to move it; global skills live under ~/.claude/skills instead of the project's .claude folder.

Gotchas

MCP servers added without --scope land in local scope and will not appear in other projects.
Unused MCP tools still load into context (his example: a set of Zapier tools). Disable the ones a session does not need.
A newly created skill showed 'unknown skill' until the session was restarted.
Superpowers makes Claude Code noticeably slower and more expensive; he only recommends it on the Max plan.
If Claude tries the gh CLI and it is missing, tell it explicitly to use the GitHub MCP server.

← Back to News digest · See also: Claude Cowork guide