Skip to main content

The Tool Selection Guide: The Hierarchy of Automation

You’ve learned the parts. You know how to build a Skill, configure a Subagent, and connect an MCP server.

But the most common question from founders and developers isn’t "How do I build X?"—it’s "WHAT should I build?"

"I want to automate my code reviews. Is that a Skill? Or a Subagent?" "I need to search my database. Do I need an MCP server, or can a Skill do that?"

This guide is your compass. We’ll cut through the complexity with a simple decision framework.

The 5-Second Decision Guide​

If you stop reading here, memorize this table. It solves 90% of architectural debates.

If you need...Then use...The "Digital FTE" Analogy
Persistent ContextCLAUDE.mdThe Employee Handbook
Repeated ProceduresSkillA Standard Operating Procedure (SOP)
Delegation & FocusSubagentA Specialized Department / Team
External Systems/DataMCPIT Access / Permitted Software
Quality Gates/SafetyHooksCompliance & Security Officers

The Decision Flow​

When you face a new automation challenge, ask these questions in order:

Question 1: Does it need external data or external systems?

  • Yes (Jira, database, API, website) → Use MCP
  • No (local files, local context) → Continue to Q2

Question 2: Is it a complex, multi-step workflow requiring autonomy?

  • Yes (research, analysis, report generation) → Use Subagent
  • No (specific, bounded task) → Continue to Q3

Question 3: Is it about enforcing rules or preventing bad actions?

  • Yes (validation, security, compliance) → Use Hook
  • No (helpful procedure) → Continue to Q4

Question 4: Do you explain this task the same way every time?

  • Yes (repetitive, consistent procedure) → Use Skill
  • No (context-dependent, general knowledge) → Use CLAUDE.md

The "Digital FTE" Mental Model​

To truly design "Digital Employees," stop thinking like a developer (functions, scripts) and start thinking like a Manager (roles, responsibilities).

1. CLAUDE.md: The Employee Handbook​

Every new hire gets a handbook. It tells them: "Here is our mission. Here is how we prefer things formatted. Here is who to ask for help."

Use when:

  • You want to set the baseline culture and context for every interaction

Don't use for:

  • Specific, granular tasks (too much noise)

2. Skills: Standard Operating Procedures​

When you hire an accountant, you give them a checklist: "Project Setup Checklist." "End-of-Month Protocol."

Use when:

  • You have a repeatable task (writing a test, formatting a blog post) that you want done your way

Don't use for:

  • Massive, ambiguous projects requiring 2 hours of autonomy

3. Subagents: Specialized Departments​

You don't ask the "General Receptionist" to audit your taxes. You send that to the "Accounting Department."

Use when:

  • The task requires a different mindset or set of tools
  • Example: A "Testing Subagent" needs different prompts than a "Creative Writing Subagent"

Don't use for:

  • Simple tasks (creating a whole department to "fix a typo" is bureaucracy)

4. MCP: IT Access & Software​

An employee can't check Salesforce if you don't give them a login. MCP is that login.

Use when:

  • The intelligence is locked away in a database, API, or website

Don't use for:

  • Things the model already knows (e.g., Python syntax)

5. Hooks: Compliance & Security​

The Compliance Officer doesn't do the work; they typically stop work that is dangerous. "Stop! You can't merge to main without approval."

Use when:

  • You need to enforce hard constraints or automate invisible logging

Don't use for:

  • Helpful suggestions (that's a Skill)

Common Scenarios: What Would You Pick?​

  • Scenario A: "I want Claude to check our Jira board for new tickets."

Analysis: Does it need external data? Yes (Jira is a project management tool outside your computer).

Selection: MCP. You need a Jira MCP server.

  • Scenario B: "Every time Claude writes code, I want it to follow our naming conventions consistently."

Analysis: Is it a repeated preference? Yes.

Selection: CLAUDE.md (if it's a global project rule) or Skill (if specific to one task type). Likely CLAUDE.md.

  • Scenario C: "I want to give Claude a 'Research Mode' where it browses the web, summarizes 50 pages, and writes a report without distracting me."

Analysis: Is it complex? Yes. Does it need focused autonomy? Yes.

Selection: Subagent. Create a "Researcher" subagent equipped with Browser MCP.

  • Scenario D: "Stop Claude from committing secrets/API keys to git."

Analysis: Is it safety/validation? Yes (preventing accidental leaks of passwords).

Selection: Hook. A pre-git-commit hook (a script that runs before saving changes).

The Strategic Takeaway​

Amateurs build prompts to conplete current work - vibe code the current task.

Professionals build Systems.

A "System" is simply the smart combination of these five elements. Your goal as an Agent Architect is to place the complexity in the right bucket.

See It in Action​

This framework is the decision-making engine. But what does it look like when applied?

Lesson 17 shows how Boris Cherny—the creator of Claude Code—actually uses these tools together in his daily workflow. 15-20 parallel sessions, Plan Mode discipline, specialized subagents for different tasks, Hooks automating the edge cases.

Study his patterns. They're blueprints for your own systems.