Skip to content
Claude with Skills

Advanced: Progressive Disclosure, Bundled Scripts, and Subagent Orchestration

This section is for practitioners who have completed the Basic and Intermediate sections. You already know every frontmatter field. This section is about architecture: how to structure skills that scale, how to bundle reusable scripts, and how to coordinate work across isolated subagents.

Prerequisites

You have completed the Intermediate section or you can write a SKILL.md with full frontmatter control, including context: fork, agent, allowed-tools, and user-invocable. You are comfortable with Python 3 and basic shell scripting.

What you will learn

LessonTopic
01 — Progressive disclosureSKILL.md as a navigator: keep context cost low as skill complexity grows
02 — Supporting scriptsBundle and invoke scripts from within skills using ${CLAUDE_SKILL_DIR}
03 — Skills calling subagentsUse context: fork and the agent field to spawn isolated subagents
04 — Subagents preloading skillsThe skills: field in subagent definitions for pre-injected domain knowledge
05 — Skills orchestrating skillsOne skill instructing Claude to invoke other skills in sequence
06 — Context window masteryDiscovery budget, body lifecycle, auto-compaction, and tuning knobs
07 — Hooks in skillsPreToolUse and PostToolUse hooks defined in skill frontmatter
08 — Plugins and distributionPackaging skills as a plugin with plugin.json for team distribution
CapstoneA multi-skill, subagent-backed code quality bot built from scratch

Learning outcomes

By the end of this section you will be able to:

  1. Structure skills for progressive disclosure so context cost stays low at scale — the navigator pattern keeps SKILL.md under 80 lines while making hundreds of lines of reference material available on demand.
  2. Bundle and invoke scripts from within skills using ${CLAUDE_SKILL_DIR}, the variable that always resolves to the directory containing SKILL.md regardless of installation location.
  3. Spawn isolated subagents from a skill using context: fork — understand when forked context helps, when it hurts, and what a subagent can and cannot see.
  4. Preload skills into a subagent’s startup context using the skills: field in a subagent definition file, so the subagent carries domain knowledge from the moment it starts.
  5. Chain skills together in an orchestration skill — understand which patterns require the main conversation and which can delegate to workers.
  6. Manage context window budget across many active skills — know the three stages of context loading, the discovery budget, and how to tune it with settings.
  7. Add lifecycle hooks to skills for automation — run PreToolUse and PostToolUse shell commands outside Claude’s turn.
  8. Package skills into a plugin for team distribution — the plugin directory layout, plugin.json manifest, and security restrictions.
  9. Build a complete multi-skill, subagent-backed system from scratch in the capstone project.

Key constraint to keep in mind

Subagents cannot spawn other subagents. If a skill uses context: fork, the forked subagent cannot itself use context: fork to spawn further subagents. All delegation flows must go through the main conversation. The orchestrator lives in the main session; workers live in forked subagents.

How to use this section

Each lesson directory contains an index.md explaining the concept and one or more complete, working example skills. The examples are designed to be installed and tested, not just read.

Terminal window
cp -r lesson-01-progressive-disclosure-in-practice/examples/pdf-toolkit ~/.claude/skills/

Lessons are ordered by concept dependency: lesson 01 (disclosure) informs lesson 02 (scripts), which both feed into lessons 03 and 04 (subagents).

Lessons