Claude Code Skill vs Subagent: The 6-Check Decision Rule

Picking Skill or Subagent in Claude Code by feel costs you afternoons of rewiring. I built my hero-image capability both ways to find the rule, and here are the 6 checks you can run in 30 seconds. 🎓 Want to learn Claude Code with me, live? Join the Claude Code Builder cohort on Maven → https://maven.com/genaiunplugged/claude-code-builder 📝 Note (Oct 2026): Claude also loads Skills on its own unless you set disable-model-invocation: true, so slash command vs auto-delegate is a hint, not a hard rule. A forked Skill (context: fork) now runs in the background by default; add background: false to wait for the result. Current docs → https://code.claude.com/docs/en/skills I built my hero-image capability twice, once as a Skill you watch and once as a Subagent you hand off to. Same job, two architectures. Along the way my old sticky note rule (making work goes in a Skill, judging work goes in an agent) fell apart, and I replaced it with one question that decides the whole thing. 📌 What you will learn - Why the "job word" does not tell you whether to build a Skill or a Subagent - The one question that decides it: do you want the work visible in your thread, or only the result? - Six checks, in order, that you can run in about 30 seconds - Why the model field means a per-turn override in a Skill and a permanent setting in a Subagent - How context: fork lets a Skill run as a Subagent, and how a Subagent can preload Skills - A live test: hero-image on 1 post versus 8 posts, same job, different answer ✅ The 6 checks 1. Visible work, or only the result? 2. Invocation: slash command, or auto-delegate? 3. File body: procedure, or identity? 4. Model field: per-turn override, or permanent? 5. Hybrid: context fork plus agent type? 6. Pick Skill or Subagent, out loud. 🛠 Tools mentioned Claude Code, SKILL.md files, Subagents, the context: fork flag, and my own hero-image, Inspector and viral-title-variants capabilities. 📋 Prerequisites Claude Code installed, and at least one Skill and one Subagent you already use. You will need both for the exercise at the end. 🎯 Who this is for Anyone building Skills and Subagents in Claude Code who wants a clear rule, so they stop wiring the wrong shape into a system they run every week. 📺 CHAPTERS: 0:00 Claude Skill vs Subagent Side by Side 0:31 Why Is Hero-Image a Skill? 1:34 Testing the Sticky Note Rule 4:10 Where the Rule Breaks 5:13 The One Question That Decides 6:16 Checks 1 to 3: Visibility, Invocation, Body 8:21 Check 4: The Model Field Trap 9:23 Check 5: Context Fork Hybrid 10:57 Build Hero-Image Both Ways 13:02 Your Exercise 🔗 RESOURCES Claude Code Skills docs → https://code.claude.com/docs/en/skills Claude Code Subagents docs → https://code.claude.com/docs/en/sub-agents Claude Code product page → https://www.anthropic.com/claude-code Official Claude documentation (Claude Code docs live here) → https://docs.claude.com Anthropic pricing and plan limits → https://www.anthropic.com/pricing 📥 Helpful Links: 🛠️ Check my AI Systems store → https://store.genaiunplugged.com/ ▶️ Watch the Full n8n Zero to Hero Course Playlist → https://tinyurl.com/3rw3x6hy 🌐 Subscribe to GenAI Unplugged Substack → https://genaiunplugged.substack.com/ 🎓 Learn with me on Maven → https://maven.com/genaiunplugged 🤝 Let's connect on LinkedIn → https://www.linkedin.com/in/dheerajsharma14/ 📚 Browse all my courses → https://genaiunplugged.com/ 🙌 A like is what tells YouTube to show this to the next person looking for the same fix. Subscribe if you want a new AI build every week. Tell me which Skill and which Subagent you tested, and which of the six checks flipped your answer, because I read every comment. ⚠️ DISCLAIMER This video is educational. Every setup I show uses the provider's own documented features. Nothing here bypasses billing, cracks licensing, or works around anyone's terms of service. I am not affiliated with any of the tools or providers shown, and there are no affiliate links in this description. Free tiers, model availability and rate limits change often, so check the official docs before you follow along. Your API keys are yours to protect. Never commit one to a repo or leave one on screen. Do not send confidential or client data through a third party model provider. Follow at your own risk. I accept no liability for any cost or issue that comes out of this tutorial. #ClaudeCode #AIAgents #Subagents