Agentic Setup for Claude — Context, Skills, and Tools

The Three Layers of Agentic Claude

Claude Code is not just an AI assistant. When configured properly, it becomes an autonomous agent that understands your project, executes complex workflows, and connects to your entire development infrastructure.

The difference between a basic Claude Code setup and a fully agentic one comes down to three layers:

• Context — what the agent knows about your project
• Skills — what the agent can do repeatedly and reliably
• Tools — what external systems the agent can interact with

Each layer multiplies the effectiveness of the others. Context without skills means the agent understands but cannot act. Skills without context means the agent can act but makes wrong decisions. Tools without either means raw capability with no direction.

Layer 1 — Context (CLAUDE.md)

CLAUDE.md is the foundation. It is a markdown file at the root of your project (or in ~/.claude/ for global instructions) that tells Claude everything it needs to know before you type a single prompt.

What to include:

• Tech stack and language versions
• Coding conventions (export style, comment policy, naming patterns)
• Architecture overview (monorepo structure, service boundaries)
• What to never do (destructive git commands, specific anti-patterns)
• Project-specific quirks and gotchas

What to avoid:

• Do not exceed 100 lines. If you need more, split into separate files and reference them.
• Do not repeat information the agent can discover from the codebase.
• Do not include temporary instructions. CLAUDE.md is for durable project context.

Layer 2 — Skills (SKILL.md)

Skills are reusable, documented workflows that the agent can execute from a single command. Each skill is a SKILL.md file with YAML frontmatter and markdown instructions.

A skill defines:

• Name and description: What this skill does
• Trigger: How to invoke it (slash command, keyword)
• Instructions: Step-by-step execution plan
• Constraints: What the skill must and must not do

Example skills for a typical project:

• PR Review: Read the diff, check conventions, verify tests, post structured feedback
• Deploy: Run pre-deploy checks, build, push, verify health endpoint
• Documentation: Generate API docs from code, update changelog, verify links
• Audit: Check dependencies for vulnerabilities, verify security headers, scan for secrets

The Agent Skills open standard means skills are portable. A SKILL.md written for Claude Code can inform workflows in other tools that adopt the standard.

Layer 3 — Tools (MCP Servers)

MCP (Model Context Protocol) servers connect Claude to external systems. They are configured in ~/.claude.json (or project-level .mcp.json) and expose three primitives:

• Resources: Data the agent can read (database records, API responses, file contents)
• Tools: Actions the agent can take (create GitHub issue, send Slack message, run query)
• Prompts: Predefined interaction patterns (structured templates for common tasks)

Common MCP connections:

• Database: Query your database directly for real data during development
• File system: Read and write files beyond the project directory
• GitHub: Create PRs, read issues, check CI status, review comments
• Slack: Send notifications, read channel context, post updates
• Playwright: Control a browser for testing, screenshots, and verification

Putting It Together

When all three layers are configured, the agent transforms from a code suggestion tool into an autonomous development partner:

• Context ensures every decision aligns with your project's conventions and architecture.
• Skills ensure complex workflows execute consistently without re-explanation.
• Tools ensure the agent can interact with real systems, not just generate text.

A practical example: you type "deploy the auth service". The agent:

1. Reads CLAUDE.md to understand the deployment conventions (Context)
2. Executes the deploy skill with pre-deploy checks and post-deploy verification (Skills)
3. Uses MCP to check GitHub CI status, push to the registry, and verify the health endpoint (Tools)

No prompt engineering. No manual steps. The environment does the work.

Conclusion

The agentic setup for Claude Code is not about finding the right prompt. It is about building the right environment. Context gives the agent knowledge. Skills give it capability. Tools give it reach. Together, they create a system that compounds in value with every task.