Repository Overview
Relevant source files
The following files were used as context for generating this wiki page:
This repository is a comprehensive collection of system prompts, model configurations, and operational guidelines for various advanced AI coding assistants and agentic IDE tools. The project serves as a technical reference for understanding how different AI agents—such as Google’s Antigravity, Qoder, Amp, and Lovable—are structured to interact with users, manage complex software engineering tasks, and integrate with local development environments.
The repository covers a wide spectrum of agentic behaviors, ranging from simple CLI-based response generation to complex, multi-modal workflows involving autonomous planning, implementation, and verification phases. It highlights the diverse approaches to task management, tool usage policies, and design aesthetics across the modern AI coding landscape.
Core Architecture and Agentic Modes
The repository documents various operational modes that agents transition through to complete complex tasks. These typically follow a structured lifecycle of planning, execution, and verification.
Antigravity Lifecycle
The Antigravity system utilizes a specific "Agentic Mode" designed to manage complex work without overwhelming the user. It relies on the task_boundary tool to synchronize progress with a task UI.
This diagram illustrates the Antigravity state machine where the agent moves between planning, execution, and verification.
Sources: Google/Antigravity/planning-mode.txt
Qoder Design Workflow
Qoder focuses on transforming high-level ideas into abstract design documents. It employs a specific sequence of repository detection and specialized documentation templates (Frontend, Backend, CLI, etc.).
Sources: Qoder/Quest Design.txt
Key Agent Identities and Tooling
The repository categorizes multiple AI entities, each with unique system instructions and toolsets.
| Agent | Developer | Primary Focus | Key Tools/Artifacts |
|---|---|---|---|
| Antigravity | Google DeepMind | Advanced Agentic Coding | task.md, implementation_plan.md, walkthrough.md |
| Amp | Sourcegraph | Software Engineering | codebase_search_agent, oracle, todo_write |
| Qoder | Qoder | Design and Implementation | search_memory, search_replace, get_problems |
| Lovable | Lovable | Web App Creation | lov- XML tags, Supabase integration, read-console-logs |
| Codex CLI | Open Source | Terminal-based Coding | apply_patch, update_plan |
Sources: Google/Antigravity/planning-mode.txt, Amp/claude-4-sonnet.yaml, Qoder/Quest Design.txt, Lovable/Agent Prompt.txt, Open Source prompts/Codex CLI/openai-codex-cli-system-prompt-20250820.txt
Task Management and Planning
A recurring theme across the repository is the use of structured TODO lists and plans to manage high-latency or complex operations.
Multi-Step Execution Flow
Agents like Amp and Codex CLI use dedicated tools to maintain a living plan. This ensures the agent does not lose context during long-running tasks.
This sequence shows the interaction between the agent and its internal task tracking system to provide visibility to the user.
Sources: Amp/claude-4-sonnet.yaml, Open Source prompts/Codex CLI/openai-codex-cli-system-prompt-20250820.txt
Technical Implementation Standards
The project defines strict standards for how agents should interact with the file system and external environments.
File Modification Strategies
Agents are instructed to avoid full file rewrites unless necessary, preferring targeted edits.
- Antigravity/Qoder: Use placeholders like
// ... existing code ...to indicate unchanged sections. - Codex CLI: Uses a specific
apply_patchformat with*** Begin Patchand*** End Patchheaders. - Amp: Prefers
edit_filefor partial changes andcreate_filefor full overwrites.
Sources: Google/Antigravity/planning-mode.txt, Qoder/Quest Design.txt, Open Source prompts/Codex CLI/openai-codex-cli-system-prompt-20250820.txt
Design Aesthetics and Web Development
For web-focused agents like Lovable and the Google web application workflow, there is a heavy emphasis on "Rich Aesthetics."
- Stack: Vite, React, Tailwind CSS, TypeScript.
- Design Rules: Use HSL color tokens, avoid generic colors, implement glassmorphism, and ensure responsive design.
- SEO: Mandatory implementation of meta tags, semantic HTML, and structured data (JSON-LD).
Sources: Lovable/Agent Prompt.txt, Google/Antigravity/planning-mode.txt
Environment and Security
The agents operate within specific sandboxed or CLI environments, such as the WebContainer (in-browser Node.js runtime) or local developer machines.
- Absolute Paths: Most agents (Amp, Qoder, Gemini CLI) require absolute paths for all file operations to prevent errors in different working directories.
- Security Policies: Explicit instructions prohibit logging secrets, committing API keys, or creating malicious code. Agents are often required to run "diagnostics" (linting, type-checking) after every edit.
- Tool Parallelism: A "Critical Instruction" found in many prompts (Amp, Cursor, Same.dev) mandates the use of parallel tool calls (e.g., reading 3 files simultaneously) to reduce user latency.
Sources: Cursor Prompts/Agent CLI Prompt 2025-08-07.txt, Open Source prompts/Bolt/Prompt.txt, Amp/claude-4-sonnet.yaml
Conclusion
The repository provides a modular and detailed framework for building and operating AI coding agents. By standardizing task lifecycles, implementation patterns, and design principles, it ensures that these agents can operate reliably across various development environments while maintaining high standards of code quality and user experience.