Vibecoding glossary

This glossary lives on the guides hub as a reference for the vocabulary that emerged around AI-assisted shipping. The terms are unsettled, some are recent inventions, some are older usages re-purposed. Treat this as a snapshot, not a dictionary.

What you'll learn

• The core terms that define the practice
• Adjacent vocabulary you will hear in the same conversations
• How a typical vibecoding workflow actually goes
• Where CobaltCapture sits in that workflow

Core terms

Vibecoding

The practice of building software primarily by directing an AI coding agent through natural-language prompts. The human states intent, the agent produces code, the human reviews the output, and they iterate. The defining feature is that the human is doing direction and review, not authorship. Originally a description of a specific working style, vibecoding has become shorthand for the broader pattern of AI-led shipping.

Prompt

The natural-language directive given to a coding agent. A good prompt names the intent ("add a sign-in form to the homepage"), the constraints ("match the existing button style, use the same auth pattern as /admin"), and the success criteria ("user lands on /dashboard after submit"). A bad prompt is just the intent and leaves the agent guessing on the rest.

Agent

The AI system that interprets prompts and writes or edits code. Distinct from an LLM chat assistant (which produces text suggestions). An agent has tools: it reads files, runs commands, applies patches, runs tests. Cursor, Claude Code, Lovable, Bolt, v0, Cline, Aider, Replit Agent, Zed AI, and Windsurf are all coding agents. They differ in modality (IDE vs CLI vs chat) but share the agent-tool-loop pattern.

Feedback loop

The cycle of prompt -> agent output -> human review -> next prompt. A tight feedback loop is the engine of vibecoding. The loop is fast when the feedback is structured: when the human can hand the agent a clear list of fixes in a format the agent can act on. The loop slows when feedback is vague or unstructured and the agent has to ask clarifying questions.

Vibe coder

The person doing the vibecoding. Not necessarily a software engineer in the traditional sense, designers, PMs, founders, and curious non-engineers are increasingly vibe-coding entire prototypes. The skill set is more about clear intent, fast judgment, and good feedback than syntax knowledge.

Adjacent terms

Agent-readable feedback

Product feedback structured so a coding agent can act on it directly: markdown, embedded screenshots with stable URLs, source URLs, dictated context. See what is agent-readable feedback for the full definition.

Vibe-shipping / MVPing

Slang for releasing a vibecoded prototype to the world fast. The point is to learn from real users before perfecting anything. Vibe-shipping leans hard on the agent producing something workable in days, not months. Distinct from a traditional MVP in that the code is largely agent-written.

Prompt engineering

The discipline of writing prompts that produce the output you want. Less of a thing than it used to be, modern coding agents tolerate sloppier prompts than the 2022-era models did. But specificity still matters, especially for design intent and edge cases.

Specification by example

The pattern of giving the agent a concrete example of what you want rather than describing it in abstract terms. "Make the button look like this screenshot" beats "make the button more modern." This is where structured visual feedback becomes essential, the screenshot IS the spec.

Pair programming with AI

The older, less-extreme cousin of vibecoding. The human is writing code with AI suggestions assisting. The human is still the author; the AI is a helper. Vibecoding flips this: the AI is the author and the human directs.

How the workflow actually goes

A typical vibecoding session has four phases.

Intent. You state what you want, in natural language, with as much specificity about constraints as you can spare. "Build a sign-in form that uses our existing button styles and posts to the /api/auth endpoint."

Generation. The agent produces a working version. Maybe it works. Maybe parts are wrong. You preview it.

Review. You walk through the result. Click, scroll, try to break it. Find the things that are off, the spacing, the copy, the missing validation. This is where most of the friction lives, because writing up what is wrong is slower than spotting what is wrong.

Feedback. You hand the agent a list of fixes. The format of the feedback determines how tight the loop is. A single Slack screenshot with "fix this", slow loop. A structured CobaltCapture review with five findings, each with cropped screenshot and dictated context, tight loop.

The loop runs three to ten times for a typical feature. Each turn is faster when the feedback is structured.

How CobaltCapture fits in

CobaltCapture is the tool that closes the feedback loop with structured visual feedback. Capture the broken screen, dictate the why, publish a review URL, paste it into the agent's next prompt. The loop stays minutes-long instead of breaking when feedback would otherwise take twenty minutes to type. See the feedback-loop-for-vibecoding spoke for the dedicated workflow.

Frequently asked questions

What is vibecoding?

Vibecoding is the practice of building software primarily by prompting an AI coding agent rather than writing code yourself. The human directs the agent in natural language, reviews the output, and iterates.

Who coined the term?

Andrej Karpathy popularized the term in early 2025 in a post describing how he was building with coding agents, accepting suggestions, glossing over details, and iterating on vibes rather than reading every diff. The term went viral and stuck.

Is vibecoding the same as no-code?

No. No-code platforms produce apps via visual builders without code being generated underneath. Vibecoding still produces code; the human just does not write it directly. The agent writes it from prompts and the human reviews or ignores the diffs.

What is the difference between vibecoding and pair programming with an AI?

Pair programming with an AI assumes the human is also writing code, with the AI providing suggestions. Vibecoding tips the balance: the agent writes nearly all the code and the human's primary job is direction, review, and feedback. The line between them is blurry but the spirit is different.

Do you still need to know how to code?

Helpful but not strictly required for prototype work. Code knowledge becomes essential when you need to debug, optimize, or maintain something the agent built. For shipping a v1 product to validate an idea, vibecoders increasingly do not write code themselves.

Frequently asked questions

What is vibecoding?

Vibecoding is the practice of building software primarily by prompting an AI coding agent rather than writing code yourself. The human directs the agent in natural language, reviews the output, and iterates.

Who coined the term?

Andrej Karpathy popularized the term in early 2025 in a post describing how he was building with coding agents, accepting suggestions, glossing over details, and iterating on vibes rather than reading every diff. The term went viral and stuck.

Is vibecoding the same as no-code?

No. No-code platforms produce apps via visual builders without code being generated underneath. Vibecoding still produces code; the human just does not write it directly. The agent writes it from prompts and the human reviews or ignores the diffs.

What is the difference between vibecoding and pair programming with an AI?

Pair programming with an AI assumes the human is also writing code, with the AI providing suggestions. Vibecoding tips the balance: the agent writes nearly all the code and the human's primary job is direction, review, and feedback. The line between them is blurry but the spirit is different.

Do you still need to know how to code?

Helpful but not strictly required for prototype work. Code knowledge becomes essential when you need to debug, optimize, or maintain something the agent built. For shipping a v1 product to validate an idea, vibecoders increasingly do not write code themselves.