Skills as Code

What Is a Skill?

• A reusable unit of functionality — invoked by name, like calling a function
• Written in natural language, stored as markdown files
• Takes input, performs work, produces output
• Lives in .claude/skills/ — version-controlled alongside your code

.claude/skills/
├── generate-deck/SKILL.md
├── verify/SKILL.md
├── publish/SKILL.md
└── generate-image/SKILL.md

Skills Are Functions

The mental model maps directly — skills are functions for an AI runtime.

Layered Abstractions

Skills can call other skills — just like functions calling functions.

/generate-deck
  ├── Creates draft slides
  ├── /verify         ← fact-checks all claims
  ├── /generate-image ← creates visuals
  └── /publish        ← exports, commits, pushes

• Each skill is self-contained but composable
• Higher-level skills orchestrate lower-level ones
• You build complex workflows from simple primitives

CLAUDE.md Is Your Config File

• CLAUDE.md sets the runtime environment — conventions, paths, preferences
• Like .bashrc, tsconfig.json, or pyproject.toml
• Every conversation loads it automatically
• Global ~/.claude/CLAUDE.md for user-wide defaults
• Project-level CLAUDE.md for repo-specific behavior

The config shapes the runtime. The skills define the operations.

The Interactive REPL

Using Claude interactively is a Read-Eval-Print Loop:

1. Read — you type a prompt
2. Eval — Claude interprets and executes
3. Print — results appear (files, code, output)
4. Loop — you iterate based on what you see

• Immediate feedback, fast iteration
• Prototype a workflow, then extract it into a skill
• The REPL is where skills are born

But the Language Is Fuzzy

2 + 2 → always 4
Same skill, same input → different outputs each time

• The runtime is non-deterministic
• This isn't a bug — it's a fundamental property
• You're programming a system that interprets intent

"The skill says what to do. The model decides how."

Testing Skills

If skills are code, they need testing — but testing looks different here.

• Behavioral testing — does the skill produce the right kind of output?
• Boundary testing — what happens with unusual inputs?
• Regression testing — does it still work after you change the prompt?
• Composition testing — do layered skills interact correctly?

You can't assert exact equality. You assert properties and patterns.
Like property-based testing, but for natural language.

The Implications

• Prompt engineering is programming — just in a fuzzy, intent-based language
• Skills are your codebase — they grow, refactor, and need maintenance
• CLAUDE.md is infrastructure — it defines the platform your skills run on
• Version control matters — skills evolve, and you need to track changes
• Composition is power — small skills combine into sophisticated workflows

We're not replacing programming. We're adding a new layer to it.