Skip to Content
Course content

2.5 Progressive Disclosure -- Structuring Large Plugins

Progressive Disclosure: Structuring Large Plugins

Lesson 2.5 — How to manage complexity in your plugins

As plugins grow, you face a challenge: too much information can overwhelm Claude and slow down responses. Progressive disclosure solves this by giving Claude only the information it needs, when it needs it.

The Problem with Large Plugins

Imagine a plugin with 10 skills, 15 commands, and 3 MCP servers. If Claude loads everything into context at once, it wastes tokens on irrelevant information and responses become slower and less accurate.

Context Overload

Loading all plugin files upfront means Claude spends time processing documentation it may never need. This leads to slower responses, higher costs, and less precise outputs.

Three Layers of Information

1

Metadata Layer

Plugin name, description, and a list of available skills and commands. Claude reads this first to understand what the plugin can do.

2

SKILL.md Layer

Detailed instructions for each skill. Claude reads a specific SKILL.md only when it needs to execute that workflow.

3

References Layer

API documentation, templates, and examples. Loaded on demand when a skill explicitly says to read a reference file.

How It Works in Practice

Plugin Structure
plugin/
  metadata.json          # Always loaded — plugin overview
  skills/
    crm/
      SKILL.md           # Loaded when CRM tasks arise
      references/
        api-guide.md     # Loaded when SKILL.md says to
    content/
      SKILL.md           # Loaded when content tasks arise
      references/
        seo-rules.md     # Loaded on demand

Claude starts by reading metadata.json. When a user asks about CRM, Claude loads the CRM SKILL.md. If that skill references an API guide, Claude loads it then — not before.

Think Like a Book

Metadata is the table of contents. SKILL.md is the chapter. References are the appendix. You would not read the entire appendix before finding the right chapter.

Keep metadata.json concise — just names and descriptions
Each SKILL.md should be self-contained for its workflow
Reference files are only loaded when explicitly needed
Claude reads only what is relevant to the current task

Key Takeaway

Progressive disclosure means Claude reads only what it needs, when it needs it. This keeps plugins fast, accurate, and scalable — whether you have 2 skills or 20.

Rating
0 0

There are no comments for now.

to be the first to leave a comment.