Build Skills that Claude actually uses well
Skills are reusable instructions that extend Claude's capabilities. But most Skills fail because they're too verbose, poorly structured, or untested.
This course teaches you the mental models behind effective Skill authoring through interactive simulations.
"The context window is a public good."
❌ Wrong approach: Write detailed explanations assuming Claude knows nothing.
✅ Right way: Challenge each token: "Does Claude really need this?"
Token Cost = f(relevance × necessity)
Your Skill competes with: system prompt, conversation history, other Skills, and the actual request. Every unnecessary token degrades Claude's performance.
Compare verbose vs concise instructions. See how token usage affects available context.
with pdfplumber.open("file.pdf") as pdf:"
CLAUDE IS SMART. ONLY ADD CONTEXT IT DOESN'T HAVE.
"Match specificity to task fragility."
❌ Wrong approach: Use same level of detail for all instructions.
✅ Right way: Tight guardrails for fragile ops, loose guidance for creative tasks.
Classify each task by appropriate freedom level.
FRAGILE OPERATIONS NEED TIGHT RAILS. CREATIVE TASKS NEED SPACE.
"SKILL.md is a table of contents, not the book."
❌ Wrong approach: Dump everything into SKILL.md.
✅ Right way: Overview in SKILL.md, details in referenced files loaded on-demand.
Design a BigQuery Skill structure. See how different architectures affect token usage.
Deeply nested references (3+ levels) cause partial reads. Keep references one level deep from SKILL.md.
BUNDLE COMPREHENSIVE RESOURCES. NO CONTEXT PENALTY UNTIL ACCESSED.
"Run validator → fix errors → repeat."
❌ Wrong approach: Execute complex operations in one shot.
✅ Right way: Plan → Validate → Execute → Verify
P(success) = P(plan_correct) × P(execute_correct)
Without validation: 0.8 × 0.8 = 0.64
With validation loop: 0.95 × 0.95 = 0.90
Design a form-filling workflow. Drag steps into optimal order.
VALIDATION CATCHES ERRORS EARLY. SCRIPTS SOLVE, DON'T PUNT.
"The description enables Skill selection from 100+ options."
❌ Wrong approach: "Helps with documents"
✅ Right way: "Extract text and tables from PDF files. Use when working with PDFs or document extraction."
[What it does] + [Trigger keywords] + [When to use]
Always third person. Max 1024 chars. No XML tags.
Write a description for an Excel analysis Skill. Get scored on discoverability.
Use gerund form: processing-pdfs, analyzing-spreadsheets
Lowercase, numbers, hyphens only. Max 64 chars.
SPECIFIC DESCRIPTIONS WIN. VAGUE DESCRIPTIONS LOSE.
"One question per module. No partial credit."
Why should Skills be concise?
Which task needs LOW freedom (exact instructions)?
What's wrong with 3+ levels of nested file references?
In "Plan → Validate → Execute → Verify", why does Validate come BEFORE Execute?
What makes a Skill description discoverable?