Reference & field notes
A working reference for the vocabulary and habits behind vibe coding. Skim the glossary, then dig into the field notes.
Glossary
The terms you'll keep running into, in plain language.
Building software by describing the outcome you want and letting an AI agent write the code, while you plan, review and steer.
An AI system that doesn't just answer — it takes actions in a loop (plan, act, observe) to accomplish a goal.
Command-line interface — a text terminal where you run tools like Claude Code by typing commands.
The natural-language instruction you give an agent. Good prompts state an outcome, constraints, and how 'done' is judged.
A short, concrete description of what to build — clear enough that an agent (or a person) could implement it without guessing.
The amount of text a model can consider at once. Anything outside it is effectively forgotten until you bring it back.
When a model states something false with confidence — like calling an API or package that doesn't actually exist.
Model Context Protocol — an open standard for connecting agents to external tools and data sources in a uniform way.
A database query where user input is passed as separate parameters, not string-concatenated — the standard defense against SQL injection.
The final project of a course where you ship something real end-to-end, applying everything you learned.
The cycle of reading an agent's output, giving feedback or a grade, and having it revise — how quality improves over iterations.
An isolated environment where code can run safely without touching your real system or data.
Field notes
Short, opinionated reference posts — drafts, shipping soon.
Why your first prompt should be a spec, not a sentence
draftA one-line ask leaves the agent guessing at scope, edge cases, and what 'done' means. Lead with a tight spec — outcome, constraints, and acceptance criteria — and you get usable code on the first pass instead of the third.
Reading a diff: the 60-second review that saves you
draftMost regressions hide in changes you never actually looked at. A fast, repeatable diff routine — scan the surface area, check the edges, run the tests — catches them before they reach main.
When to trust the agent and when to read every line
draftBoilerplate and well-trodden patterns rarely need scrutiny; security boundaries, money, and data migrations always do. Calibrating that judgment is the difference between fast and reckless.