Course Review and Exam Tips
The Whole Course on One Page
You've gone from 'what is a skill?' to building multi-file, subagent-executing, org-distributed skills. Let's consolidate. A skill is a folder with a SKILL.md that teaches Claude a task once, loaded on demand via progressive disclosure, following the agentskills.io open standard. Everything else builds on that.
| Topic | The essential |
|---|---|
| What & why | Folder + SKILL.md; teach once, apply automatically; open standard, write-once |
| Progressive disclosure | name+description always; full body on match; files/scripts on demand |
| Creating | Directory name = command; only description recommended |
| Frontmatter | allowed-tools, disable-model-invocation, user-invocable, context: fork, paths |
| Advanced | multi-file, !`cmd` injection, $ARGUMENTS, run-don't-read scripts |
| Scope & sharing | Enterprise > Personal > Project; git / plugins / managed |
| Skills + subagents | Subagents don't inherit skills — list them in skills:; or context: fork |
The entire course distilled into the essentials of each topic.
High-Yield Facts
- •A skill = folder + SKILL.md (frontmatter + instructions); follows agentskills.io; 'write once, use everywhere'.
- •Progressive disclosure: only name+description loaded at startup; full body on match; supporting files/scripts only when referenced.
- •The DIRECTORY name is the command (/deploy-staging); frontmatter name is just the display label. Only description is recommended.
- •Key frontmatter: allowed-tools (pre-approve), disable-model-invocation (you-only), user-invocable: false (Claude-only), context: fork (run as subagent), paths (limit auto-activation).
- •Scripts run without loading their source; !`cmd` injects live command output before Claude sees the skill; $ARGUMENTS/$N parameterize.
- •Scope precedence: Enterprise > Personal > Project; plugins namespaced. Share via git (team), plugins (community), managed settings (org).
- •Subagents don't inherit skills — list them in the subagent's skills: field; a skill can run in a subagent with context: fork.
Two questions answer most skill decisions
(1) 'Should this be in context always, or only when relevant?' → CLAUDE.md vs a skill. (2) 'Who should trigger this — me, Claude, or both?' → disable-model-invocation / user-invocable. Master those two and the rest follows.
Common Traps
- 1.Thinking the frontmatter 'name' sets the command — it's the DIRECTORY name (except a plugin-root SKILL.md).
- 2.Believing allowed-tools restricts tools — it PRE-APPROVES; restrict with disallowed-tools or permissions.
- 3.Forgetting subagents don't inherit skills — you must list them in the skills: field.
- 4.Putting always-on facts in a skill instead of CLAUDE.md.
- 5.Vague or near-duplicate descriptions — the top cause of non-triggering and wrong-skill selection.
- 6.Reading a script into context instead of running it — wastes the progressive-disclosure benefit.
- 7.Confusing context: fork (skill runs AS a subagent) with the skills: field (subagent CARRIES skills).
Watch the loading boundary
Most skill mistakes come from misunderstanding what's loaded and when: only name+description at startup, full body on match, files/scripts on demand, and (in subagents) skills preloaded in full. Design every skill around that loading model.
Where to Go Next
You now have a complete mental model of Agent Skills — from a single SKILL.md to org-wide distribution. To cement it, build something real:
- •Turn your most-repeated instruction into a personal skill with a specific, keyword-rich description; test that it triggers.
- •Build a multi-file skill with a bundled script you tell Claude to RUN, referenced via ${CLAUDE_SKILL_DIR}.
- •Add !`cmd` dynamic context injection to ground a skill in live data (e.g. !`git diff HEAD`).
- •Commit a project skill to .claude/skills for your team, then try wiring it into a custom subagent's skills: field.
- •Read the official skills docs and the agentskills.io standard to reinforce the wire-level details.
You've completed the course
From progressive disclosure to multi-file skills, dynamic injection, subagent execution, and org-wide sharing, you can now create, configure, share, and troubleshoot Agent Skills — and pick the right Claude Code feature every time. Congratulations.
Key Takeaways
- ✓A skill is a folder + SKILL.md that teaches Claude a task once; it loads on demand via progressive disclosure and follows the agentskills.io 'write once, use everywhere' standard.
- ✓Progressive disclosure: name+description always in context, full body on match, supporting files/scripts only when referenced.
- ✓The directory name is the command; only description is recommended; key frontmatter = allowed-tools, disable-model-invocation, user-invocable, context: fork, paths.
- ✓Two questions decide most things: 'in context always vs only when relevant?' (CLAUDE.md vs skill) and 'who should trigger it?' (invocation control).
- ✓Scope precedence is Enterprise > Personal > Project (plugins namespaced); share via git, plugins, or managed settings.
- ✓Subagents don't inherit skills — list them in the skills: field, or run a skill as a subagent with context: fork; avoid vague descriptions and run (don't read) bundled scripts.
Check Your Understanding
Test what you learned in this lesson.
Q1.What sets the command you type to invoke a skill?
Q2.Which statement about allowed-tools is correct?
Q3.What is true about skills and subagents?
Q4.Which two questions resolve most skill design decisions?
Practice This Lesson