Commands vs Skills vs Agents in Claude Code — What Goes Where

Configure Claude Code so it knows your stack, follows your conventions, runs repeatable workflows,...

---

title: Commands vs Skills vs Agents in Claude Code — What Goes Where

published: true

date: 2026-06-05 14:19:03 UTC

tags: anthropicclaude,claudecode,claudecowork,claudeai

canonical_url:

---

![](https://cdn-images-1.medium.com/max/1024/1*l4sCiR0ib-rNjYcQ8SphDw.png)

Configure Claude Code so it knows your stack, follows your conventions, runs repeatable workflows, and delegates to specialists — without repeating yourself every session.

What you’ll build

A production-style Claude Code project layout:

1. CLAUDE.md — team instructions (committed)

2. CLAUDE.local.md — personal overrides (gitignored)

3. .claude/settings.json — permissions and environment (committed)

4. .claude/rules/ — modular instruction files

5. .claude/skills/ — slash commands and auto-invoked workflows

6. .claude/agents/ — isolated subagent personas

Everything Claude needs about your project lives in one place — commit .claude/ to git so the whole team shares it.

Tool stack

markdown

| Tool | Role |
|------|------|
| **Claude Code** | CLI agent with tools, memory, skills, subagents |
| **`CLAUDE.md`** | Project memory loaded at session start |
| **`.claude/settings.json`** | Permissions, hooks, env vars |
| **Skills** | Reusable prompts — manual `/name` or automatic |
| **Agents** | Focused sub-sessions with their own tools |
| **Rules** | Path-scoped or global instruction modules |

Layers

markdown

| Layer | Always on? | Trigger |
|-------|------------|---------|
| `CLAUDE.md` + `rules/` | Yes — every session | Automatic |
| `settings.json` | Yes — gates tool use | Automatic |
| Skills | On demand or auto | `/project:name` or model decides |
| Agents | On demand | User delegates or Claude spawns |
| Hooks | Yes — around tool calls | `settings.json` → `hooks` |

Session workflow

1. Developer runs claude in a configured repo

2. Memory + rules + permissions load automatically

3. Skills and agents handle specialized work on demand

4. Team shares the same .claude/ tree via git

Prerequisites

markdown

| Requirement | Check |
|-------------|--------|
| Claude Code installed | `claude --version` |
| A git repository | `git status` |
| Terminal access to your project | — |

Install Claude Code: [code.claude.com](https://code.claude.com/docs/en/overview).

Part 1 — Understand the layout

#### Project root

plaintext

your-project/
├── CLAUDE.md # Team instructions (committed)
├── CLAUDE.local.md # Personal overrides (gitignored)
└── .claude/
    ├── settings.json # Permissions + config (committed)
    ├── settings.local.json # Personal permissions (gitignored)
    ├── rules/ # Modular instruction files
    ├── skills/ # Workflows with SKILL.md
    ├── commands/ # Legacy single-file skills (optional)
    └── agents/ # Subagent definitions

#### Global home directory

Claude also reads ~/.claude/ (all projects):

plaintext

~/.claude/
├── CLAUDE.md # Your global defaults
├── settings.json # Global permissions
├── skills/
├── age

Leer artículo completo en dev.to

// artículos relacionados

Twitter/X: 🚀 While you’re sleeping, the market is printing. I’m a futures trader, sharing my personal high…

Twitter/X15 jun

Twitter/X: anthropic is literally crushing the new model benchmark https://t.co/Mhyy2SZtUQ

Twitter/X15 jun

Twitter/X: EU regulator evaluating implications of Anthropic Mythos curbs after US directive https://t.co/FE…

Twitter/X15 jun