You describe the outcome in natural language and an AI model writes, edits, and explains the code. You stay in the loop - accepting, rejecting, and refining - but you're directing, not typing every line. Coined by Andrej Karpathy, it's how most people now start building.
You hand a goal to an autonomous agent that plans, writes code, runs commands, installs packages, and fixes its own errors across many steps with little or no human approval in between. More powerful, and more dangerous, because the review steps that catch mistakes are skipped.
Plain-English glossary
The 10 terms every new vibe coder should know
- Prompt
- The instruction you give the AI. Better prompts - specific, with examples and constraints - produce dramatically better code.
- Model
- The underlying AI (e.g. Claude, GPT, Gemini). More capable models reason better about architecture and catch more of their own mistakes.
- Context window
- How much of your project the AI can "see" at once. When it forgets earlier decisions, you've usually exceeded it.
- Agent
- An AI that acts on its own across many steps - writing files, running commands, installing packages - toward a goal you set.
- MCP (Model Context Protocol)
- A standard that lets agents safely connect to tools and data (databases, APIs, your filesystem) instead of guessing.
- Hallucination
- When the AI confidently invents something false - a function, an API, or a package - that does not actually exist.
- Dependency / package
- Pre-built code your app pulls in (from npm, PyPI, etc.). AI loves adding them, sometimes ones that are fake or malicious.
- Environment variable / secret
- Sensitive config like API keys. The #1 beginner mistake is committing these into your repo where anyone can read them.
- Deploy
- Putting your app on the public internet. "Runs locally" and "safe in production" are different milestones.
- Static analysis
- Checking code for problems without running it - the fast, deterministic way to catch whole classes of bugs before users do.
Ready to ship what you build?
PeakStack reviews AI-built code for security, scalability, and cost on every commit - with the exact file, line, and fix.
Explore PeakStack