Configuration and Usage Patterns

Relevant source files

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

This page documents the configuration mechanisms, access methods, and operational patterns for Amp. It covers VSCode settings, Thread YAML inspection, model selection, and environment context structure. For information about Amp's tool ecosystem, see Tool Suite and Capabilities.

---

VSCode Configuration Settings

Amp operates as a VSCode extension and requires configuration through the editor's user settings. The primary configuration parameters control the service endpoint and model selection.

Core Settings

Setting	Type	Purpose	Default
amp.url	string	Amp service endpoint URL	https://ampcode.com/
amp.gpt5	boolean	Enable GPT-5 model selection	false

Model Selection Configuration

Amp supports multiple underlying LLM models with model-specific system prompts. The amp.gpt5 flag switches between configuration profiles:

• amp.gpt5: false - Uses Claude 4 Sonnet configuration with extensive reasoning capabilities
• amp.gpt5: true - Uses GPT-5 configuration optimized for parallel execution and minimal reasoning output

VSCode Settings Configuration

Sources: Amp/README.md12-20

---

Thread YAML Access Method

Amp maintains conversation state in Thread objects accessible through a YAML inspection interface. This mechanism allows users to view the complete system prompt, environment context, and tool definitions for debugging and transparency.

Access Procedure

1. Execute any query in Amp to create an active Thread
2. Hold Alt (Windows) or Option (macOS)
3. Click the workspace button in the Amp interface
4. Select "View Thread YAML" from the menu

The Thread YAML contains:

• System prompt - Complete instructions for the LLM
• Environment context - Working directory, OS, repository info
• Tool definitions - Full schema for all available tools
• Conversation history - Previous messages and tool calls
• Debug information - Token usage, model parameters

Sources: Amp/README.md1-9 Amp/view-thread-yaml.png1

---

Environment Context Structure

Each Thread includes a structured environment section that provides the LLM with contextual awareness of the execution environment. This section is automatically populated and included in the system prompt.

Environment Data Fields

Field	Example Value	Purpose
Today's date	Mon Sep 15 2025	Temporal context for queries
Working directory	/c:/Users/ghuntley/code/system-prompts-and-models-of-ai-tools	Current operation location
Workspace root folder	/c:/Users/ghuntley/code/system-prompts-and-models-of-ai-tools	Project root for absolute paths
Operating system	windows (Microsoft Windows 11 Pro 10.0.26100)	Platform-specific command syntax
Repository	https://github.com/ghuntley/system-prompts-and-models-of-ai-tools	Source control context
Amp Thread URL	https://ampcode.com/threads/T-...	Thread identifier
Directory listing	<directoryListing>...</directoryListing>	Cached workspace structure

Directory Listing Cache

The <directoryListing> section provides a cached snapshot of the workspace structure, reducing the need for explicit list_directory tool calls:

c:/Users/ghuntley/code/system-prompts-and-models-of-ai-tools
├ .git/
├ .github/
├ Amp/
├ Claude Code/
├ Lovable/
├ Qoder/
├ Windsurf/
└ README.md

Sources: Amp/gpt-5.yaml514-612 Amp/claude-4-sonnet.yaml756-852

---

Model-Specific System Prompt Differences

The two supported models receive tailored system prompts optimized for their respective strengths. The prompts share core functionality but differ in emphasis and verbosity constraints.

GPT-5 Configuration Characteristics

Emphasis on Concision

• "MINIMIZE REASONING" directive - Limit verbose reasoning blocks throughout session
• Brief 1-2 sentence summaries before tool calls
• Direct action without verbose validation

Parallel Execution Policy

• Default to parallel for all independent work: reads, searches, diagnostics, writes, subagents
• Serialize only for strict dependencies

Output Constraints

• Keep visible output under 4 lines unless detail requested
• Never pad with meta commentary
• Avoid inner monologue

Claude 4 Sonnet Configuration Characteristics

Emphasis on Agency

• "Do the task end to end" - Complete resolution of requests
• Balance initiative with restraint
• No additional code explanation unless requested

Oracle Integration

• Use Oracle tool FREQUENTLY for planning, review, analysis, debugging
• Mention to user why Oracle is invoked
• Language: "I'm going to ask the oracle for advice"

Task Management

• Use todo_write and todo_read VERY frequently
• Mark todos as completed immediately
• Do not batch multiple tasks

Sources: Amp/gpt-5.yaml63-70 Amp/gpt-5.yaml72-107 Amp/claude-4-sonnet.yaml10-58 Amp/claude-4-sonnet.yaml211-286 Amp/claude-4-sonnet.yaml301-356

---

Common Usage Patterns

Both configurations share core usage patterns for tool execution, file operations, and workflow management, with differences in how aggressively they are applied.

Tool Execution Pattern

Sequential Discovery

1. Use broad search tools (codebase_search_agent, Grep, glob)
2. Read relevant files with Read tool
3. Make changes with edit_file or create_file
4. Verify with get_diagnostics and test commands

Parallel Execution (emphasized in GPT-5)

• Multiple independent Read calls in single turn
• Parallel codebase_search_agent queries for different concepts
• Concurrent Task tool invocations for disjoint file modifications

File Operation Workflow

Read Before Edit

• Always use Read tool to understand file content before editing
• Check imports and surrounding context
• Identify existing patterns and conventions

Edit Strategy

• Use edit_file for targeted string replacements
• Use create_file for new files or complete rewrites
• Set replace_all: true for repeated pattern changes

Verification Gate

• Run get_diagnostics after modifications
• Execute lint commands from AGENTS.md if present
• Run typecheck/build commands

Subagent Delegation Pattern

Decision Matrix

Need	Tool	When to Use
Senior engineering advice	oracle	Planning, review, architecture, debugging
Conceptual code search	codebase_search_agent	Find logic by description
Heavy multi-file execution	Task	Feature scaffolding, refactors, migrations

Workflow Example

1. oracle(plan feature) → Strategic guidance
2. codebase_search_agent(validate scope) → Locate existing patterns
3. Task(execute) → Implement across files

Sources: Amp/gpt-5.yaml220-298 Amp/claude-4-sonnet.yaml35-68 Amp/claude-4-sonnet.yaml211-297

---

AGENTS.md Auto-Context Integration

Amp automatically includes the AGENTS.md (or legacy AGENT.md) file in the system prompt if present in the workspace. This mechanism allows project-specific conventions to be persistently available without manual injection.

AGENTS.md Content Structure

Common Commands

• Typecheck: npm run typecheck, cargo check, go build
• Lint: npm run lint, eslint .
• Build: npm run build, cargo build
• Test: npm test, pytest, go test ./...

Code Style Preferences

• Naming conventions (camelCase, snake_case, PascalCase)
• Import ordering rules
• Error handling patterns
• Testing approach and framework

Project Structure

• Directory organization
• Module boundaries
• Key file locations

Usage Pattern

When Amp needs to verify code changes:

1. Check for commands in AGENTS.md
2. Execute typecheck/lint/build commands from file
3. If commands unknown, ask user and suggest appending to AGENTS.md

Sources: Amp/gpt-5.yaml300-314 Amp/claude-4-sonnet.yaml455-471

---

Configuration Access Pattern Summary

The complete configuration access workflow combines VSCode settings, Thread inspection, and runtime environment context.

Key Configuration Points

Layer	Configuration Entity	Access Method
VSCode	amp.gpt5, amp.url	User settings JSON
Thread	System prompt, environment	Thread YAML viewer (Alt/Option + Click)
Workspace	Commands, conventions	AGENTS.md auto-inclusion
Runtime	Working directory, OS	Automatic environment detection

Sources: Amp/README.md1-20 Amp/gpt-5.yaml1-612 Amp/claude-4-sonnet.yaml1-852