May 23, 2026 10 min EN claude-code cli mcp sub-agents plugins skills

Understanding Claude Code in 10 minutes

What Claude Code is

Claude Code is an Anthropic CLI tool that lets you delegate coding tasks to Claude from a terminal. It reads/writes files, executes shell commands, and orchestrates sub-agents. Source: Anthropic — “Claude Code lets developers delegate coding tasks to Claude directly from their terminal”.

The foundational building blocks

1. The context (context window)

Claude works within a limited window. Everything read, written, or returned by a tool consumes this window. Beyond 80%, automatic compaction kicks in — but Claude is often already drifting before that threshold. Source: .claude/rules/session-practices.md.

Practical rule: never exceed 50% and expect good results.

2. Tools

Claude can call predefined tools: Read, Write, Edit, Bash, Glob, Grep, Task (sub-agent), etc. Each tool call is a structured exchange (JSON Schema).

3. Sub-agents

A sub-agent is a Claude instance launched by the main agent with its own context. The sub-agent does its work and then returns only the summary. The main context remains intact.

This is the primary mechanism for exploring a large codebase without saturating the session.

4. Slash-commands

Text commands prefixed with /. Native examples: /clear (reset session), /compact (summarize the context), /rewind (roll back to the last clean checkpoint).

ulk adds: /ulk:bruce, /ulk:godspeed, etc. — each corresponding to a markdown file in framework/agents/.

5. Skills

A skill is a folder with a SKILL.md describing a capability (creating a PPTX, auditing accessibility, deploying to Vercel). Claude reads the skill before executing the task to align on best practices.

ulk bundles third-party skills: Figma (7), Swift (7), Flutter (2), context-audit (1), cwb-app-icon (1, opt-in). Source: CLAUDE.md section “Community Skills”.

6. Plugins

Installable bundles (skills + commands + MCP). ulk relies on official Anthropic plugins (/feature-dev, /simplify, /commit, /pr-review-toolkit:review-pr, /frontend-design, /claude-md-improver).

Reference: framework/agents/_shared/plugins-protocol.md.

7. MCP (Model Context Protocol)

Standard protocol for connecting Claude to external sources: Figma, Linear, Notion, Cloudflare, Vercel, etc. An MCP server exposes tools that Claude can call.

ulk prefers CLI when available (0 network tokens, local execution) and falls back to MCP. Source: CLAUDE.md section “CLI Tools” — “Priority rule: CLI available → use it (0 tokens). No CLI + MCP configured → use MCP.”

The 4 context hygiene rules

Canonical reference: framework/agents/_shared/context-hygiene-protocol.md.

Rule	When to apply
1. `/rewind` instead of correcting	When Claude goes in the wrong direction. The failed attempt pollutes everything that follows.
2. `/clear` when switching tasks	New task = new session. Persist state externally (issue, todo.md) before clearing.
3. Sub-agents for heavy exploration	Delegation = clean context, returns only the summary.
4. Proactive `/compact` at 50-60%	With explicit instructions (preserve X, forget Y), before the automatic compact at 80%.

Recommended workflow

START — read the task from issue/todo.
RESEARCH (sub-agent) — explore the code.
PLAN (sub-agent) — propose an approach.
IMPLEMENT — execute, test, commit.
REVIEW — update issue/todo, /clear.

Source: .claude/rules/session-practices.md.

Persistent memory

Claude Code can maintain memory between sessions via:

CLAUDE.md at the project root — repo instructions (loaded at every session).
MEMORY.md + individual files in memory/ — semantic memory (user, feedback, project, reference).

ulk extends this with an Obsidian-compatible Knowledge Vault Loop (agent Lovecraft 47), which automatically captures/distributes/surfaces memory between sessions. Source: CLAUDE.md section “Knowledge Vault Loop”.

TL;DR

Claude Code = CLI + tools + sub-agents + slash-commands + skills + plugins + MCP. ulk adds 90 pre-built markdown agents, shared protocols, and hygiene rules to prevent context saturation.