Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.replit.com/llms.txt

Use this file to discover all available pages before exploring further.

AI tools help you give instructions in natural language and get useful output back. On Replit, that output can become plans, code, designs, explanations, debugging steps, and changes to your project. You do not need to understand AI deeply before building. A few concepts help you give better instructions and evaluate the results with confidence.

​
Large language models

Most AI builders are powered by large language models, often called LLMs. An LLM is a model trained to recognize patterns in text and generate text that follows those patterns. Text can mean prose, code, HTML, CSS, SQL, JSON, terminal commands, test output, error messages, or configuration. That is why the same kind of model can help write copy, explain a bug, create code, or summarize a plan.

​
Text generation

An LLM does not retrieve a fixed answer from a database. It generates a response piece by piece based on the input you provide and the context it can see. When you write a prompt, the model uses your words, the conversation, and any available project context to predict a useful next response. That response might be an explanation, a plan, or text that another system can use to take action. This means the quality of your input matters. Clear goals, examples, constraints, and context make good results more likely.

​
Tokens

AI models process text in small units called tokens. A token can be a word, part of a word, punctuation, whitespace, or a piece of code. Tokens matter because they affect how much information an AI model can consider at once. Long conversations, large files, logs, screenshots, and instructions all contribute to the context the model has to reason from. You do not need to count tokens while building. You do need to understand that Agent works better when the important information is clear, current, and focused.

​
AI can vary

AI output is not always identical. The same request can produce different wording, designs, implementation details, or tradeoffs. That variability is useful for exploration, but it also means you should review important work. Test what Agent builds, ask for explanations when something is unclear, and use checkpoints when you need a safe recovery point.

​
What makes AI output better

Agent performs best when you provide:
  • A clear goal
  • The audience or user need
  • Constraints and non-goals
  • Examples, screenshots, files, data, or links
  • The current problem or exact error
  • Acceptance criteria for what “done” means
Compare these two prompts:
  • “Write code for a website.”
  • “Write code for a website to collect email addresses for my personal mailing list. Store submissions in Google Sheets.”
The second prompt names the goal, the data, and the destination, so Agent has enough to act without guessing. This information is called context. To learn why context matters, see Context management.

​
From AI to AI agents

A chat model can respond to a message. An AI agent uses a model plus tools to perform multi-step work — it can reason about a goal, choose tools, inspect information, make changes, check results, and continue until the task is done or needs your input.

​
Models and tools

An agent has two important parts:
PartWhat it does
Language modelUnderstands instructions, generates text, reasons about options, and decides what to do next.
ToolsLet the agent take action, such as reading files, editing code, running commands, searching docs, testing an app, or publishing changes.
The model decides what action might help. The tools let the agent act on that decision.

​
What agents can do

Agents are useful for work that takes more than one step, such as:
  • Planning a feature
  • Creating files and code
  • Reading existing project structure
  • Running commands or tests
  • Debugging errors
  • Updating designs
  • Connecting services
  • Summarizing what changed
The agent does not just answer; it works through a task.

​
Replit Agent

Replit Agent is an AI agent for building on Replit. It can help you create apps and other artifacts, inspect your project, edit files, run commands, test behavior, explain code, create tasks, and guide publishing workflows. For example, you can ask Agent to: 
Build a simple event signup app for a community meetup.
Visitors should see event details, submit their name and email, and get a confirmation message.
Keep the first version simple so I can publish it today.
Agent can turn that request into a project, create a first version, and help you refine it.

​
You still lead the work

Agent can move quickly, but you still make the important decisions:
  • What the artifact should do
  • Who it is for
  • What should be in scope
  • What should not change
  • Whether the result works
  • When it is ready to publish
This is why context management and review matter. Agent is strongest when you give it clear direction and test the result.

​
Common misconceptions

  • “AI should know what I mean.” AI works from the information you give it. If something matters, say it clearly.
  • “The first answer should be final.” Building with AI is iterative. Review, test, and refine.
  • “AI understands intent the way people do.” AI can infer patterns, but it does not automatically know your taste, audience, constraints, or definition of done.
  • “An agent is just a chatbot.” A chatbot mainly responds. An agent can use tools and take action.
  • “Agent knows the whole product goal automatically.” Agent works from your instructions and context. If something matters, include it.
  • “Agent removes the need to review.” Agent can build, but you decide whether the artifact matches your intent.

​
Where to go next

Context management

Learn how context windows shape AI output.

Vibe coding

Put AI, context, agents, and apps together into a builder workflow.

Build with Agent

Learn practical habits for collaborating with Replit Agent.

Agent

Learn more about Agent’s product capabilities.