Campus AI Development Vibe coding

The continuum Band 01

01

Vibe coding

Chat → run → prompt again. You read little or none of the code.

Also called: Prompt-driven development (the more structured form)

01Definition

The person describes what they want in plain language, runs what the AI produces, and reacts to the result with another prompt. Code goes largely unread; the running output is what gets judged. Prompt-driven development is the more structured form: the human still breaks the work into a deliberate sequence of prompts.

02Origins

Andrej Karpathy coined the term in February 2025 for a style where you "fully give in to the vibes" and forget the code exists, which he framed as fine for throwaway weekend projects.

Within weeks the term spread well beyond developers. Simon Willison argued in March 2025 that if you have reviewed, tested and understood the code, it isn’t vibe coding; it’s software development with AI help. Collins named “vibe coding” its Word of the Year for 2025.

By February 2026 Karpathy wrote that agent-driven programming had become a professional default “except with more oversight and scrutiny,” and proposed the name agentic engineering for that. Vibe coding kept its original meaning: building without reading the code.

03How it works

Person AI
01Describe the goal
02Generate the app
03Run it and look
04Paste back errors

Steps 2–4 repeat until it looks right.

  1. 01Describe the goal in plain language.
  2. 02Run what the AI produces, often in a browser-based builder or chat window.
  3. 03Paste back errors, or describe what looks wrong.
  4. 04Repeat until the result looks right. The code itself is rarely opened.

Typical tools: chat assistants; app builders such as Replit, Lovable and Bolt; coding agents set to accept every change.

04Profile across the dimensions

How the work is done

Human role
Describes and reacts
Unit of work
A conversation
Source of truth
The chat history
Agent autonomy
Varies: often run in agentic tools that edit files and run commands; what defines it is not reading the output
Lifecycle coverage
Build only

How you know it’s right

Code read by a person
Little or none
Verification
Run it and look
Traceability
None
Delivery automation needed
None

Fit and risk

Upfront investment
Minutes
Durability
Throwaway
Data and risk ceiling
Public or sandbox data only

Compare all five bands

05Lifecycle coverage

Plan
Design
Build
Test
Deploy
Operate

Solid: AI does core work · Hatched: partial or informal · Dashed: done by people. Compare all bands

06When to use it

Good for: Discovery, learning, demos, throwaway tools

One personWhere it works best: personal prototypes and one-off tools.
A teamBarely works. Nobody else understands the code, and two people vibe-coding one app overwrite each other. The usual pattern is that one person prototypes, then the team rebuilds at band 03 or 04.

Avoid it for

  • Anything other people will depend on
  • Personal, student or other institutional data
  • Sign-in, payments or access control
  • Anything you couldn’t throw away tomorrow

Signs you’ve outgrown it

  • Someone else starts using it
  • It touches real data
  • You can’t explain how it works
  • Bugs come back after they’ve been fixed

07Development environment

Where it runs: A browser-based app builder (Replit, Lovable, Bolt), a chat assistant’s preview pane, or a personal laptop.

Data
Fake or public data only
Credentials
None that reach real systems
Hosting
A personal sandbox, never a campus domain
On campus
A sanctioned sandbox account under the enterprise agreement, so prompts stay inside contract terms

Compare environments across bands

08Minimum controls

  • Personal space or sandbox only
  • No real or institutional data
  • Throw it away, or rebuild it properly before sharing

Risks

  • Security holes: hard-coded credentials, missing access checks and unvalidated input, none of which show up when you only look at the output.
  • Comprehension debt: no one understands the code well enough to fix it.
  • Data leakage: pasting institutional data into consumer tools outside campus agreements.
  • Shadow IT: a prototype quietly becomes a service no one supports.

09On campus

10Sources

  1. Karpathy, post on X introducing “vibe coding,” Feb. 2, 2025
  2. Willison, “Not all AI-assisted programming is vibe coding,” Mar. 2025
  3. Collins Dictionary, Word of the Year 2025
  4. Karpathy on “agentic engineering,” Feb. 2026, and Sequoia AI Ascent talk, Apr. 2026
  5. Sarigoz, “Vibe coding vs spec-driven dev: a per-task decision guide,” BizStack, Sept. 11, 2026
  6. Sarkar & Drosos, “Vibe coding: programming through conversation with artificial intelligence,” arXiv, 2025