Context★

Context (also called prompt context or session context) is the total text payload compiled by your application harness and sent in a model provider request. It is the model's active "working memory" for that specific run.

You build and structure context whenever you:

• Write system instructions defining an agent's persona.
• Implement RAG pipelines to insert document search results into a prompt.
• Serialize local code files so an assistant can read them.

Technical Details: In-Context Learning

Because models are stateless and their parameters are frozen, they cannot learn new facts permanently during a chat session. Instead, they perform In-Context Learning:

1. Reading Comprehension: The model processes the injected files and instructions in its attention heads, using them to guide its predictions.
2. Context Assembly: A standard agent harness compiles multiple data sources into a structured string before sending the request:

   System Instructions ──► Chat History ──► Retrieved Documents ──► User Message = Context Block
   

3. Vibe vs. Fact: While parameters dictate how a model thinks (grammar, code style, reasoning), the context dictates what it thinks about (your database schema, today's date, your project specs).

Field Applications & Formatting

To prevent the model from confusing system rules with user files, engineers use XML tags or clear Markdown headers to separate context areas:

1. Fullstack Developers (Context Tagging)

• Prompt Template Example:

  You are an automated code audit assistant. Analyze the source code provided in the tags below for security bugs.
  
  <source_code file="auth.ts">
  ${fileContent}
  </source_code>
  
  Output your audit report as a bulleted list.