What is a prompt?

A prompt is the instruction a builder gives an AI model: a request, a question, or a description of what the model should produce. In vibe coding, prompts replace much of what specifications, tickets, and code used to do. The prompt is the primary building tool, and everything the model produces traces back to what the prompt said, implied, or left out.

Prompts range from a single line ("add a dark mode toggle") to structured briefs with goals, constraints, and examples. The model treats all of it as context for predicting what output would satisfy the request.

What makes a prompt effective?

Specific beats vague. "Add a login form with email and password fields, inline validation, and clear error messages" gives the model a testable target; "make login better" leaves it guessing at intent. Effective prompts name the outcome, the constraints, and the things that must not change.

Context does the rest of the work. The model only knows its training data plus whatever the conversation and visible files contain, so a good prompt supplies what's missing: what the app does, what already exists, and what just changed. Examples are especially powerful, and showing the model one correct output usually beats describing it in the abstract.

One goal per prompt keeps results debuggable. When a prompt bundles 5 requests and the output is wrong, there's no way to tell which instruction failed.

How do prompts work in a vibe coding session?

A session is a chain of prompts. The first one sets the goal, and follow-ups refine, correct, and extend: "the button is misaligned on mobile," "use the same style as the settings page," "that broke search, undo it." Error messages pasted into the chat act as prompts too, often the most effective ones, since they tell the model exactly what failed.

The chain accumulates in the model's context window, which is why long sessions drift. Summarizing decisions, or starting a fresh session with a short recap, keeps early instructions from getting lost.

How does a prompt differ from a specification?

A specification defines the whole system upfront; prompts build incrementally, one request at a time. That looseness makes vibe coding fast, and it's also the main source of surprises: whatever the prompt didn't say, the model decides. Builders working on anything beyond a quick prototype often keep a short written spec and paste it into new sessions, giving the prompt chain a stable source of truth.

What prompt patterns work when building software?

A few shapes come up constantly in vibe coding. The plan-first prompt asks the model to propose an approach before touching anything ("describe how this would be implemented before writing code"), which surfaces misunderstandings while they're still cheap. The constraint prompt states what must not change ("don't modify the database schema"), protecting working parts from enthusiastic rewrites.

The example-driven prompt shows one correct instance and asks for more of the same, ideal for repetitive structures like form fields or card layouts. The recap prompt restates the project's goal and key decisions at the top of a fresh session, resetting a drifting conversation. And the audit prompt turns the model on its own work: "list what could break on bad input" or "check this for exposed secrets."

How has prompting changed as models improved?

Early language models rewarded incantations: magic phrases, role-play setups, and step-by-step scaffolding. Modern models need less ritual and more information. The craft has shifted from phrasing tricks toward supplying the right material, a practice increasingly called context engineering. Clear intent, concrete constraints, and relevant examples now matter far more than exact wording.