The complete foundation for building software through AI prompting — what it is, when it works, and how to do it well.

## Key Points

1. **Describe** what you want (a prompt)
2. **Review** what the AI generated
3. **Test** whether it actually works
4. **Refine** by describing what to fix or change
5. **Commit** when a unit of work is stable
- Keep prompts focused on one change at a time
- Test after every generation, not after five
- Use version control as your undo button
- Do not try to describe an entire application in one prompt
- If you spend more than 3 iterations on the same issue, stop and diagnose manually
- **Prompt stacking:** Giving 5 instructions at once, getting 3 right and 2 wrong, then trying to fix the 2 while preserving the 3. This creates a tangle.
- **Review skipping:** Accepting generated code without reading it. This compounds — bad code in round 1 becomes the foundation for round 2.

Vibe Coding Fundamentals

The complete foundation for building software through AI prompting — what it is, when it works, and how to do it well.

---

What Vibe Coding Actually Is

Vibe coding is writing software by describing what you want in natural language and letting an AI tool generate the code. You guide the direction, review the output, and iterate. The AI writes most of the code; you steer.

This is not "no-code." You still need to read code, understand what it does, and catch when the AI goes wrong. The skill ceiling is high. The difference from traditional development is that your primary output is prompts and reviews, not keystrokes.

The core mental model: You are the architect and the QA team. The AI is the construction crew. You do not need to lay every brick, but you need to know if the building is sound.

---

The Prompting Loop

Every vibe coding session follows the same cycle:

1. Describe what you want (a prompt)
2. Review what the AI generated
3. Test whether it actually works
4. Refine by describing what to fix or change
5. Commit when a unit of work is stable

This loop runs dozens or hundreds of times per session. The speed of the loop is the speed of your development.

Making the Loop Fast

• Keep prompts focused on one change at a time
• Test after every generation, not after five
• Use version control as your undo button
• Do not try to describe an entire application in one prompt
• If you spend more than 3 iterations on the same issue, stop and diagnose manually

Common Loop Failures

• Prompt stacking: Giving 5 instructions at once, getting 3 right and 2 wrong, then trying to fix the 2 while preserving the 3. This creates a tangle.
• Review skipping: Accepting generated code without reading it. This compounds — bad code in round 1 becomes the foundation for round 2.
• Test deferral: "I'll test it all at the end." By then you have 15 interleaved changes and no idea which one broke things.

---

When Vibe Coding Works Well

Strong Use Cases

Prototypes and MVPs: Speed matters more than perfection. You want to validate an idea, not ship production code. Vibe coding can compress weeks into hours.

CRUD applications: Standard create-read-update-delete apps with a database, API, and frontend. AI models have seen millions of these. The patterns are well-established.

Internal tools and dashboards: Lower stakes, fewer users, less need for optimization. Admin panels, data viewers, reporting tools.

Landing pages and marketing sites: Static or near-static content with standard layouts. AI excels at generating these.

Scripts and automation: One-off scripts, data transformations, file processing, API integrations. Well-defined inputs and outputs.

Learning and exploration: Trying a new framework or language. The AI can generate idiomatic starter code faster than reading documentation.

Why These Work

• Well-established patterns in training data
• Clear success criteria (it works or it does not)
• Low cost of bugs (easy to spot, easy to fix)
• Limited scope (one app, one purpose)

---

When Vibe Coding Fails

Weak Use Cases

Distributed systems: Consensus, partitioning, eventual consistency. AI will generate code that looks right but has subtle race conditions and data loss bugs.

Real-time systems: Audio/video processing, game engines, trading systems. Latency-sensitive code requires understanding hardware and timing that AI cannot reason about reliably.

Safety-critical software: Medical devices, automotive, aviation. The review burden exceeds the generation benefit. You need formal verification, not vibes.

Cryptography and security primitives: Never let AI write your auth logic from scratch. Use established libraries. AI will confidently produce insecure implementations.

Large-scale refactors of existing codebases: AI loses context. It will fix one file and break the three files that depend on it.

Performance-critical hot paths: AI generates correct but slow code. It favors readability over performance and does not understand your specific bottlenecks.

Why These Fail

• Bugs are invisible until production (or worse, until an audit)
• Correctness requires reasoning about global state across many components
• The cost of a bug is catastrophic, not just inconvenient
• Patterns are rare in training data or require domain expertise

---

Context Management

The context window is the AI's short-term memory. Everything it can see — your prompt, the files, the conversation history — must fit in this window. Managing it is a core vibe coding skill.

Practical Context Strategies

Keep files small. A 500-line file is easier for the AI to work with than a 2000-line file. If a file grows past 300 lines, consider splitting it.

Use project instruction files. Claude Code uses CLAUDE.md, Cursor uses .cursorrules, Copilot uses .github/copilot-instructions.md. These files persist context across sessions. Put your conventions, architecture decisions, and constraints there.

Reference files explicitly. Do not assume the AI remembers a file from 20 messages ago. Say "In src/auth/login.ts, change the validation logic" rather than "change the validation logic."

Prune conversation history. Long conversations accumulate stale context. After 30-40 exchanges, start a new session with a fresh summary of where you are.

Front-load important information. Put constraints and requirements at the top of your prompt, not buried in paragraph three.

Context Anti-Patterns

• Pasting entire log files into the prompt (give the relevant 10 lines, not 500)
• Asking the AI to "remember" something from earlier (it has no persistent memory beyond the window)
• Including irrelevant files in the context hoping the AI will figure out which ones matter
• Never updating your project instruction file (it should evolve with the project)

---

When to Take Manual Control

Vibe coding is not an all-or-nothing approach. Knowing when to type code yourself is a critical skill.

Take Control When

You know exactly what to change. If the fix is a one-line change and you know which line, just make the edit. Prompting the AI to do it takes longer than doing it.

The AI is looping. If you have described the same fix three times and the AI keeps generating variations that do not work, it does not understand the problem. Read the code yourself, diagnose, and either fix it or write a much more specific prompt.

You need to understand the code. Sometimes you need to read and trace through code manually to understand what is happening. This is not a failure of vibe coding; it is part of the process.

The change spans many files. AI tools work best on single-file or few-file changes. If you need coordinated changes across 10 files, plan the change manually and have the AI execute each piece.

You are debugging a subtle issue. Step-through debugging, adding console logs, checking network requests — these are manual activities. Use the AI to hypothesize causes, but do the investigation yourself.

Signs You Should Not Take Control

• You are about to write boilerplate (let the AI do it)
• You are writing code in an unfamiliar language (let the AI draft, then review)
• You are tempted to copy-paste from Stack Overflow (prompt the AI instead; it will adapt the solution to your context)

---

The Right Mindset

Vibe coding requires a different mindset than traditional development:

Speed over perfection, then perfection over speed. In the first phase, move fast and accept imperfect code. In the second phase, slow down and clean up what matters.

Read more than you write. Your job is reviewing, not typing. If you are not reading the generated code carefully, you are accumulating debt you cannot see.

Think in iterations, not specifications. You will not get it right in one prompt. Plan to iterate. Each round should make the code better, not try to make it perfect.

Trust but verify. The AI will be confidently wrong sometimes. It will use deprecated APIs, invent function signatures, and produce code that compiles but does not do what you asked. Always test.

Maintain your own understanding. If you cannot explain what the code does, you cannot maintain it. Do not let the AI generate code you do not understand.

---

Quick-Start Checklist

For your first vibe coding session:

1. Pick a simple project (a CRUD app, a landing page, a CLI tool)
2. Set up version control before you start
3. Write a 2-3 sentence description of what you want to build
4. Choose one AI tool and stick with it for the session
5. Commit after every working change
6. Review every generated file before moving on
7. Stop and reflect after 1 hour — what worked, what did you struggle with?

---

Anti-Patterns Summary

Anti-Pattern	Why It Hurts	Fix
One giant prompt	AI loses focus, generates partial results	Break into sequential prompts
Never reading generated code	Invisible bugs compound	Review every generation
Using vibe coding for everything	Some problems need manual expertise	Know the boundaries
Ignoring version control	No undo when AI breaks things	Commit constantly
Fighting the AI for 10 rounds	Diminishing returns after 3 attempts	Diagnose manually, write targeted prompt
No project instruction file	AI re-learns your conventions every session	Maintain CLAUDE.md or equivalent
Huge context, no focus	AI gets confused by irrelevant information	Curate what the AI sees