Skip to content

Claude Code Concepts

A visual guide to Claude Code features and when to use them. Adapted for Seer Agent Engine from luongnv89/claude-howto.

Already know Claude Code?

Skip to Feature Decision Flowchart or Quick Reference.

Feature Decision Flowchart

Use this flowchart to decide which feature to use for your needs:

flowchart TD
    A["What do you need?"] --> B{"Persistent<br/>context?"}
    B -->|Yes| C{"Shared with<br/>team?"}
    B -->|No| D{"User-triggered<br/>or automatic?"}

    C -->|Yes| E["CLAUDE.md / AGENTS.md<br/>Git-tracked project context"]
    C -->|No| F["~/.claude/CLAUDE.md<br/>Personal preferences"]

    D -->|User triggers| G{"External data<br/>needed?"}
    D -->|Automatic| H{"React to events<br/>or context?"}

    G -->|Yes| I["MCP Tool<br/>API integration"]
    G -->|No| J["Command<br/>/content-audit"]

    H -->|Events| K["Hook<br/>PostToolUse, Stop"]
    H -->|Context| L["Skill<br/>Auto-loaded behavior"]

    M{"Parallel or<br/>isolated work?"} --> N["Subagent<br/>Separate context"]
    D -->|Heavy task| M

    O{"Share across<br/>team?"}
    J --> O
    L --> O
    K --> O
    O -->|Yes| P["Plugin<br/>Bundle & distribute"]

    style E fill:#e1f5ff
    style F fill:#e1f5ff
    style I fill:#fff9c4
    style J fill:#c8e6c9
    style K fill:#ffccbc
    style L fill:#e1bee7
    style N fill:#ffe0b2
    style P fill:#d1c4e9

Quick Reference Table

Need Use Example Seer Implementation
Persistent project context CLAUDE.md Architecture, standards AGENTS.md in repo root
User-triggered workflow Command Run audit, generate report /content-audit, /search-landscape
Automatic behavior Skill Apply standards automatically writing-standards, seo-methods
React to events Hook Validate after edits, QA at stop PostToolUse, Stop in core-dependencies
External data access MCP Tool Query BigQuery, fetch SERP data BigQuery, DataForSEO servers
Parallel isolated work Subagent Review + test simultaneously Code reviewer, test engineer
Share with team Plugin Distribute workflows seo@seer-agent-core

Feature Comparison Matrix

Category Skill Command Hook MCP Subagent
Triggered By Agent User Event Both Both
Context Efficiency High High N/A Low High
Parallelizable - - - - Yes
Team Shareable Yes Yes Yes Config Yes
Can Use Other Features Yes Yes Limited N/A Yes
Token Cost Low Medium None Medium Isolated

Key Distinctions

Skills vs Commands

graph LR
    subgraph Skills ["Skills (Automatic)"]
        S1["Agent decides when to load"]
        S2["Based on context/task"]
        S3["Always-on behavior"]
    end

    subgraph Commands ["Commands (Manual)"]
        C1["User explicitly invokes"]
        C2["Triggered by /command"]
        C3["One-time workflow"]
    end

    S1 --> S2 --> S3
    C1 --> C2 --> C3

Seer Examples:

Type Name When It Activates
Skill writing-standards Automatically when creating client deliverables
Skill seo-methods Automatically for any SEO-related task
Command /content-audit Only when you type /content-audit
Command /chain-of-thought Only when you want step-by-step reasoning

Hooks vs Skills

Aspect Hook Skill
Trigger Lifecycle event (deterministic) Task context (reasoning)
Purpose Validate, track, enforce Guide, advise, behave
Example "Run linter after every edit" "Use Seer writing voice"

Principle: Hooks for certainty, Skills for thinking.

Commands vs Skills for Resources

Critical Limitation

Commands cannot have resource folders. Only skills support progressive disclosure (loading resources/ on demand).

Feature Skill Command
Can have resources/ folder Yes No
Progressive loading Yes No
Reference materials In resources/ Must use skills

If a command needs substantial reference material, create a skill and reference it:

# In command frontmatter
---
name: content-audit
skills: [seo-methods]  # Skill contains methodology resources
---

The Activation Flow

How Seer Agent Engine uses hooks to activate skills automatically:

sequenceDiagram
    participant User
    participant Hook as UserPromptSubmit Hook
    participant Engine as Activation Engine
    participant Skills as Skill Registry
    participant Claude

    User->>Hook: "Create a content audit for this page"
    Hook->>Engine: Scan prompt
    Engine->>Skills: Match keywords & patterns
    Skills-->>Engine: writing-standards, seo-methods
    Engine->>Claude: Inject: "Consider using these skills..."
    Claude->>Claude: Load relevant skills
    Claude->>User: Generate audit with standards applied

The three lifecycle hooks in core-dependencies:

Hook When Purpose
UserPromptSubmit Before response Inject skill suggestions
PostToolUse After Edit/Write Track file changes
Stop Before completion QA checkpoint

Practical Examples

Example 1: SEO Content Audit

You type: /content-audit

What happens:

  1. Command triggers workflow
  2. UserPromptSubmit hook suggests skills
  3. seo-methods skill loads (analysis techniques)
  4. writing-standards skill loads (Seer voice)
  5. Workflow executes with both behaviors active
  6. Stop hook runs QA checkpoint

Example 2: Strategic Recommendation

You type: "Write recommendations for improving organic traffic"

What happens:

  1. UserPromptSubmit hook matches "recommendations" + "organic"
  2. Suggests seo-methods and writing-standards
  3. Claude loads skills automatically
  4. Response uses Seer voice and SEO methodology
  5. No command needed - skills activated by context

Example 3: Using MCP for Data

You type: "Get ranking data for seerinteractive.com"

What happens:

  1. Claude recognizes need for external data
  2. Calls BigQuery MCP tool
  3. Fetches ranking data from Seer Signals
  4. Returns structured results
  5. Skills can then analyze/present the data

Plugin Architecture Overview

How plugins bundle features in Seer Agent Engine:

graph TB
    subgraph core["core-dependencies (Required)"]
        H["Hooks"]
        SS["Shared Skills"]
        CC["Core Commands"]
    end

    subgraph seo["seo plugin"]
        SC["SEO Commands"]
        SM["seo-methods Skill"]
    end

    subgraph innovation["innovation plugin"]
        IC["Agent Commands"]
        IS["Building Skills"]
    end

    core --> seo
    core --> innovation

    H --> |activates| SM
    SS --> |standards| SC

Plugin lineup:

Plugin Purpose Required?
core-dependencies Hooks, shared skills, activation Yes
seo SEO workflows No
analytics Analytics workflows No
pdm Paid media workflows No
innovation Agent building, R&D No
operations Time tracking, Wrike No

When to Use What

Decision Scenarios

Scenario Use Why
"Always apply Seer voice to deliverables" Skill Automatic, context-driven
"Run a content audit workflow" Command User-triggered, explicit
"Check code quality after every edit" Hook Event-driven, deterministic
"Query BigQuery for rankings" MCP Tool External data access
"Review code AND run tests in parallel" Subagents Isolated, parallel contexts
"Share SEO workflows with team" Plugin Bundled distribution
"Remember project architecture" CLAUDE.md Persistent context

Anti-Patterns

Don't Do Instead
Create command for one-time task Just ask Claude directly
Build skill for user-triggered workflow Create a command
Put shared logic in each plugin Centralize in core-dependencies
Hardcode credentials Use environment variables

Deep Dive Resources

Official Documentation

External Guides

Seer Agent Engine Docs

Quick Answers

What's the difference between a skill and a command?
  • Skill = Agent loads automatically based on context
  • Command = You invoke it explicitly (/content-audit)
When should I create a hook vs a skill?
  • Hook = Deterministic reaction to events (PostToolUse, Stop)
  • Skill = Context-aware behavior guidance
Can commands have reference materials?

No. Only skills support resources/ folders. If your command needs substantial references, create a skill and reference it in the command's frontmatter.

Why is core-dependencies required?

It provides the activation engine (hooks) and shared standards. Without it, skills won't auto-activate and standards won't be centralized.

How do I know which skills are active?

Claude tells you which skills it's using. You can also run /core-doctor to see activation status.

See Also