System Overview and LLM Configuration

Relevant source files

• Amp/README.md
• Amp/claude-4-sonnet.yaml
• Amp/gpt-5.yaml
• Amp/view-thread-yaml.png

Purpose and Scope

This page documents Amp's core system architecture, identity, and LLM configuration mechanisms. It covers the multi-model support system, VSCode configuration settings, and system prompt retrieval methods. For detailed information about specific sub-systems, see Oracle AI Advisor Architecture, Sub-agent Delegation and Task Tool, Codebase Search Agent, and Tool Suite and Capabilities.

System Identity and Core Principles

Amp is a powerful AI coding agent built by Sourcegraph that assists users with software engineering tasks including adding functionality, solving bugs, refactoring code, and explaining code. The system operates with a defined agency model that balances initiative with restraint.

Core Behavioral Principles

Amp follows three primary behavioral mandates defined in its system prompt:

1. End-to-End Task Completion: "Do the task end to end. Don't hand back half-baked work. FULLY resolve the user's request and objective." The system continues working through problems until reaching complete solutions rather than stopping at partial answers.

2. Initiative-Restraint Balance: When users request plans, Amp provides plans without immediately editing files. When users request implementation, Amp takes action. "If the user asks for a plan, give a plan; don't edit files."

3. Concise Communication: "Do not add explanations unless asked. After edits, stop." The system minimizes verbose output and avoids unnecessary preamble or postamble.

Sources: Amp/gpt-5.yaml10-32 Amp/claude-4-sonnet.yaml5-33

Multi-Model Architecture

Amp implements a configurable LLM backend system that supports multiple language models with model-specific system prompts optimized for each provider's capabilities.

Model Configuration System

Model Selection and System Prompt Routing

Sources: Amp/README.md12-20 Amp/gpt-5.yaml1-13 Amp/claude-4-sonnet.yaml1-10

Model-Specific Optimizations

Each model receives a tailored system prompt optimized for its specific capabilities:

Model	System Prompt File	Key Optimizations
Claude 4 Sonnet	claude-4-sonnet.yaml	Default configuration; optimized for tool use patterns and concise communication
GPT-5	gpt-5.yaml	Includes reasoning minimization directives: "MINIMIZE REASONING: Avoid verbose reasoning blocks"

The GPT-5 prompt includes specific instructions not present in the Sonnet prompt: "Before any significant tool call, state a brief summary in 1-2 sentences maximum. Keep all reasoning, planning, and explanatory text to an absolute minimum."

Sources: Amp/gpt-5.yaml63-69 Amp/claude-4-sonnet.yaml1-753

Configuration System

VSCode Settings Schema

Amp configuration is managed through VSCode user settings with two primary configuration keys:

VSCode Configuration to System Behavior Mapping

Configuration example:

• amp.url: Specifies the Amp service endpoint URL (default: "https://ampcode.com/")
• amp.gpt5: Boolean flag enabling GPT-5 model usage (default: false, uses Claude 4 Sonnet)

Sources: Amp/README.md13-20

System Prompt Retrieval Mechanism

Amp exposes its active system prompt through a Thread YAML view accessible via keyboard shortcut interaction:

System Prompt Retrieval Flow

Retrieval Steps:

1. Login to Amp via VSCode
2. Issue any query to create an active thread
3. Hold Alt (Windows) or Option (macOS)
4. Click the workspace button
5. Select "View Thread YAML" option
6. System prompt appears in YAML structure under system key

Sources: Amp/README.md1-10 Amp/view-thread-yaml.png1

System Prompt Architecture

Core System Prompt Structure

Both model-specific system prompts follow a common architectural structure organized into major sections:

System Prompt Section Hierarchy

Sources: Amp/gpt-5.yaml14-542 Amp/claude-4-sonnet.yaml10-753

Key Architectural Sections

Role and Agency Definition

The system prompt establishes Amp's identity and operational mandate in the opening section. For GPT-5, this includes: "You are Amp, a powerful AI coding agent built by Sourcegraph. You help the user with software engineering tasks."

Key directives:

• "Do the task end to end. Don't hand back half-baked work."
• "Balance initiative with restraint"
• "Do not add explanations unless asked. After edits, stop."

Sources: Amp/gpt-5.yaml10-32 Amp/claude-4-sonnet.yaml5-33

Guardrails Section

Establishes four core constraints:

• Simple-first: "prefer the smallest, local fix over a cross-file 'architecture change'"
• Reuse-first: "search for existing patterns; mirror naming, error handling, I/O, typing, tests"
• No surprise edits: "if changes affect >3 files or multiple subsystems, show a short plan first"
• No new deps: "without explicit user approval"

Sources: Amp/gpt-5.yaml31-44

Fast Context Understanding

Defines discovery methodology:

1. "In parallel, start broad, then fan out to focused subqueries"
2. "Deduplicate paths and cache; don't repeat queries"
3. "Avoid serial per-file grep"
4. Early stop conditions: when exact files/symbols are identified or bug locus is found

Sources: Amp/gpt-5.yaml46-61

Parallel Execution Policy

Mandates default parallel execution for independent work: "Default to parallel for all independent work: reads, searches, diagnostics, writes and subagents."

Parallelization applies to:

• Reads/Searches/Diagnostics
• codebase_search_agent tool calls for different concepts
• oracle tool for distinct concerns
• Task tool executors with disjoint write targets

Serialization required when:

• Planning must complete before dependent code edits
• Write conflicts exist (same file or shared contracts)
• Chained transforms where step B requires artifacts from step A

Sources: Amp/gpt-5.yaml72-122

AGENTS.md Auto-Context System

Both prompts reference an automatic context injection system: "This file (plus the legacy AGENT.md variant) is always added to the assistant's context."

The AGENTS.md file documents:

• Common commands (typecheck, lint, build, test)
• Code-style and naming preferences
• Overall project structure

Sources: Amp/gpt-5.yaml300-314 Amp/claude-4-sonnet.yaml456-470

Environment Context Injection

The system prompt includes dynamic environment context populated at runtime:

Runtime Environment Data Injection

Example environment block structure:

• Today's date: "Mon Sep 15 2025"
• Working directory: Full absolute path
• Workspace root folder: Project root path
• Operating system: "windows (Microsoft Windows 11 Pro 10.0.26100 N/A Build 26100) on x64"
• Repository: GitHub repository URL
• Amp Thread URL: Direct link to current thread
• Directory listing: Cached tree structure of workspace

Sources: Amp/gpt-5.yaml514-612 Amp/claude-4-sonnet.yaml755-851

Tool System Integration

Tool Registration Architecture

The system prompt includes complete tool definitions in a structured schema format. Each tool is defined with:

• type: function
• name: Tool identifier used in function calls
• description: Usage instructions and guidelines
• parameters: JSON schema defining input structure

Amp Tool Ecosystem Organization

Sources: Amp/gpt-5.yaml625-1462 Amp/claude-4-sonnet.yaml867-2462

Tool Communication Protocol

The system prompt establishes strict rules for tool usage:

• "ALWAYS follow the tool call schema exactly as specified"
• "NEVER refer to tool names when speaking to the USER"
• "Instead, just say what the tool is doing in natural language"
• "If you need additional information that you can get via tool calls, prefer that over asking the user"

Sources: Amp/gpt-5.yaml140-158 Amp/claude-4-sonnet.yaml872-983

todo_write and todo_read Integration

Amp includes dedicated task management tools with specific usage mandates:

Tool Purpose: "You plan with a todo list. Track your progress and steps and render them to the user."

Usage Frequency: "Use these tools frequently to ensure that you are tracking your tasks and giving the user visibility into your progress."

Completion Discipline: "MARK todos as completed as soon as you are done with a task. Do not batch up multiple tasks before marking them as completed."

Sources: Amp/gpt-5.yaml160-218 Amp/claude-4-sonnet.yaml301-399

Communication and Output Formatting

Markdown Formatting Rules

The system prompt enforces strict markdown conventions:

Rule	Requirement
Bullets	Use hyphens - only
Numbered lists	Only for procedural steps
Headings	#, ##, ###; don't skip levels
Code fences	Always add language tag; no indentation
Inline code	Wrap in backticks; escape as needed
Links	Every file name must be file:// link
Style	No emojis, minimal exclamation points

Sources: Amp/gpt-5.yaml354-376

File Linking Syntax

Amp mandates specific file reference formatting using the file:// URI scheme:

file:///absolute/path/to/file.ext
file:///absolute/path/to/file.ext#L32
file:///absolute/path/to/file.ext#L32-L42

Special character handling: "Always URL-encode special characters in file paths (spaces become %20, parentheses become %28 and %29, etc.)."

Example: file:///Users/alice/My%20Project%20%28v2%29/test%20file.js

Sources: Amp/gpt-5.yaml372-399 Amp/claude-4-sonnet.yaml554-641

Concise Communication Mandate

Both prompts enforce brevity with specific token minimization goals:

GPT-5 Specific: "MINIMIZE REASONING: Avoid verbose reasoning blocks throughout the entire session. Think efficiently and act quickly."

Claude 4 Sonnet Specific: "IMPORTANT: Keep your responses short. You MUST answer concisely with fewer than 4 lines (excluding tool use or code generation), unless user asks for detail."

Examples from prompt:

• User: "4 + 4" → Response: "8"
• User: "How do I check CPU usage on Linux?" → Response: "top"

Sources: Amp/gpt-5.yaml63-69 Amp/claude-4-sonnet.yaml668-743

Verification and Quality Systems

Verification Gate Sequence

The system prompt mandates a specific verification order: "Typecheck → Lint → Tests → Build"

Gate Execution Rules:

• Use commands from AGENTS.md or discover via repository search if unknown
• Report evidence concisely with counts and pass/fail status
• If pre-existing unrelated failures block progress, document scope limitations

Sources: Amp/gpt-5.yaml332-344

Quality Bar Requirements

Code quality mandates:

• Match style of recent code in same subsystem
• Small, cohesive diffs; prefer single file if viable
• Strong typing, explicit error paths, predictable I/O
• No as any or linter suppression unless explicitly requested
• Add/adjust minimal tests if adjacent coverage exists
• Reuse existing interfaces/schemas; don't duplicate

Sources: Amp/gpt-5.yaml316-330