Prompt Analysis

prompts.md

Analysis of AI Assistant System Prompts

Analysis Overview

The provided prompts define a diverse set of AI assistants, each with a unique persona, toolset, and operational paradigm. They can be broadly categorized into two groups:

1. Primary Agent Prompts: These define the core identity and functionality of different AI coding assistants (e.g., Cascade, v0, Devin, Roo). They specify how the agent should interact with the user, edit code, and use tools.
2. Meta-Task & Supporting Prompts: These are specialized prompts or data files that either support a primary agent (like a Tools.json file) or define a meta-task for the AI, such as evaluating the quality of a potential "memory."

I will first analyze the Primary Agent Prompts, followed by the Meta-Task Prompts, and conclude with a summary of key themes and differences.

Table 1: Primary Agent Prompts Analysis

This table breaks down the core AI assistants defined in the prompts.

Prompt File Name	AI Persona / Name	Core Purpose & Context	Key Characteristics & Paradigm
Prompt-windsurf.txt	Cascade (by Codeium)	Agentic pair programming in a general IDE.	AI Flow Paradigm: Works independently and collaboratively. Memory System: Proactively uses create_memory to persist context. Code Edits: Combines all changes into a single edit_file call using {{ ... }}. Tooling: Emphasizes concise, non-redundant tool use and has specific tools like browser_preview and deploy_webb_app.
Prompt-vscode.txt	GitHub Copilot	AI programming assistant within the VS Code environment.	IDE Integration: Tailored for VS Code, with tools like get_vscode_api. Code Edits: Uses insert_edit_into_file with // ...existing code... comments. Validation: Explicitly instructed to call get_errors after editing a file. Style: Short, impersonal, and follows Microsoft content policies.
Prompt-v0.txt	v0 (by Vercel)	AI-powered assistant for generating UI components, primarily for Next.js and React.	Generative UI: Focuses on creating front-end code from scratch. MDX Output: Responses are in MDX format, using custom components like <CodeProject> and <QuickEdit>. Framework-Specific: Heavily optimized for a browser-based Next.js runtime with shadcn/ui and Tailwind CSS pre-installed.
Prompt-trae.txt	Trae AI	Agentic pair programming within a specific IDE, responding to user tasks and thoughts.	Instructional & Reflective: Receives a user task and a "thought" on it, then decides if a tool is needed. Limited Toolset: The prompt explicitly states no tools are available, forcing direct responses. Editing Style: Uses // ... existing code ... for placeholders.
Prompt-samedev.txt	Unnamed (by Same)	AI coding assistant for a cloud-based IDE (Same.new) focused on web development and UI cloning.	Website Cloning: Explicit instructions and tools (web_scrape) for pixel-perfect UI cloning. Development Workflow: Uses Bun over npm, auto-configures Vite/Next.js, and uses a versioning tool. Editing Style: Uses // ... existing code ... <description> comments.
Prompt-replit.txt	Unnamed (by Replit)	Expert autonomous programmer building software on the Replit platform.	Platform-Specific: Tightly integrated with Replit's environment (workflows, packager, database tools). Iterative Feedback: Relies on feedback tools (web_application_feedback_tool) to ask for user confirmation. Editing Tool: Uses a unique str_replace_editor tool with its own command set.
Prompt-bolt.txt	Bolt	Expert AI assistant for the WebContainer in-browser runtime.	Environment-Constrained: Aware of its sandboxed environment (no pip, no C++ compiler). Database-Focused: Defaults to Supabase with extremely strict rules for SQL migrations. Artifact-Based Output: Generates a single <boltArtifact> containing all file and shell actions.
Prompt-cline.txt	Cline	Highly skilled software engineer operating in a terminal-based agentic environment.	Iterative Tool Use: Strictly one tool per message, waiting for the result before proceeding. Plan vs. Act Mode: Explicitly defines two operational modes for planning and execution. Editing Style: Uses a replace_in_file tool with a unique <<<<<<< SEARCH/=======/>>>>>>> REPLACE format.
Prompt-codex.txt	Unnamed (Codex CLI)	Terminal-based agentic coding assistant in a git-backed workspace.	Git-Aware: Works within a git repository and is aware of git log and pre-commit hooks. Patch-Based Edits: Uses an apply_patch command with a specific diff format. High Autonomy: Instructed to keep working until the query is completely resolved.
Prompt-roo.txt	Roo	Highly skilled software engineer focused on maintainability and minimal code changes.	Surgical Edits: Features a precise apply_diff tool requiring start/end line numbers. Mode Switching: Can request to switch between modes (e.g., code to architect) via the switch_mode tool. Multi-Modal: Has defined modes for Code, Architect, Ask, Debug, etc.
Prompt-lovable.txt	Lovable	AI editor for web apps with a live iframe preview.	Live Preview Context: Aware that the user can see live changes. Component-Focused: Follows atomic design principles, creating small (< 50 lines) components. Custom XML Commands: Uses tags like <lov-write> and <lov-code> to structure its entire response.
Prompt-Junie.txt	Junie	A helpful assistant for exploring project structures in a readonly mode.	Readonly Analysis: Explicitly forbidden from modifying files. Its purpose is to investigate and answer questions. Specialized Search: Uses search_project for fuzzy search and get_file_structure to understand layouts. Terminal-like Interaction: Responds with <THOUGHT> and <COMMAND> tags.
Prompt-dia.txt	Dia (by The Browser Co)	An AI chat product inside the Dia web browser, focused on providing rich, decorated answers.	Rich Content Generation: Not a coding agent. Generates text with "Simple Answers" (<strong>), images (<dia:image>), and custom hyperlinks (ask://ask/...). Strict Formatting Rules: Detailed rules about when and where to place images.
Prompt.txt	Devin	A real code-wiz software engineer using a real computer OS.	Planning & Standard Modes: Has a "planning" mode for information gathering and a "standard" mode for execution. Must call <suggest_plan> before acting. Comprehensive Tooling: Has a broad set of tools, including shell, editor, LSP, browser, and deployment commands. Environment Awareness: Reports environment issues rather than trying to fix them.
Agent Prompt.txt	Unnamed (Cursor AI)	A Claude 3.7 Sonnet-powered agentic assistant in the Cursor IDE.	Parallel Tool Use: Explicitly encouraged to call multiple tools in parallel for efficiency. Standardized Edits: Uses a generic edit_file tool. Context-Driven: Receives rich context from the IDE (open files, cursor position, linter errors).

Table 2: Meta-Task & Supporting Prompts Analysis

These prompts define secondary or supporting roles for the AI system.

Prompt File Name	Purpose & Context	Key Characteristics & Paradigm
Memory Prompt.txt	To judge whether a "memory" captured from a user-AI conversation is worth remembering.	AI as a Judge: Sets up an AI to act as a quality filter. Scoring System: Uses a 1-5 scoring system with specific criteria (general applicability vs. task-specific detail). Strong Negative Bias: Explicitly told to "err on the side of rating things POORLY."
Memory Rating Prompt.txt	To extract potentially useful "memories" from a conversation for later evaluation.	Information Extraction: Its goal is to identify and formulate a potential memory from a conversation. Strict Criteria: Provides detailed positive and negative criteria for what constitutes a good memory. Structured Output: Must return the potential memory in a specific JSON format.
Agent loop.txt / Modules.txt / Prompt-manus.txt	Defines the operational structure and capabilities for an agent named Manus.	Modular Architecture: Describes a system composed of a Planner, Knowledge, and Datasource module. Event-Driven: Operates on a chronological event stream (Message, Action, Observation). Structured Planning: Follows numbered pseudocode plans provided by the Planner module. Authoritative Data: Prioritizes using Datasource APIs over general web search.

Key Themes and Comparative Analysis

This collection of prompts reveals several key trends and differing philosophies in the design of agentic AI systems.

1. Persona and Branding: Each prompt gives the AI a distinct identity, from the professional "GitHub Copilot" to the creative "v0" and the helpful "Lovable." This branding is tied to the platform it operates on (Microsoft, Vercel, Replit), creating a cohesive user experience.

2. Code Editing Mechanisms: This is the area with the most diversity, reflecting different trade-offs between precision, ease of use, and model intelligence.

   • Placeholder-Based ({{...}}, //...): Used by Cascade, Copilot, and SameDev. This is a concise way to specify changes but relies on a "smarter" apply model to correctly place the edits.
   • Diff/Patch-Based (<<<<<, apply_patch): Used by Cline and Codex. This is more explicit and less ambiguous than placeholders but can be more verbose.
   • Surgical/Line-Based (apply_diff): Used by Roo. Requiring start/end line numbers is the most precise method, minimizing ambiguity but requiring more upfront analysis by the agent.
   • Command-Based (str_replace_editor): Used by Replit. This abstracts file editing into a stateful tool, mimicking a command-line editor.
   • Full-File Rewrite (<lov-write>, <boltAction>): Used by Lovable and Bolt. This is the safest way to avoid ambiguity but is also the most verbose and potentially costly.

3. Interaction and Autonomy Models:

   • Turn-by-Turn: Most assistants (like Copilot or Roo) operate in a classic request-response loop.
   • Planning Modes: Cline and Devin have explicit "planning" modes, where they first gather information and get user buy-in on a high-level plan before executing. This is a powerful pattern for complex tasks.
   • Agentic Loops: Manus and Codex are designed to run in a continuous loop, taking actions until the task is complete, minimizing user interaction.

4. Environment and Platform Integration:

   • Many prompts are deeply tied to a specific environment (Replit, VS Code, Vercel v0, WebContainer). This allows them to use platform-specific tools and be aware of environmental constraints (e.g., Bolt's "no pip" rule). This specialization leads to more powerful and reliable behavior within that context.

5. Meta-Cognition and Self-Improvement:

   • The "Memory" prompts for Cascade and the "Thinking" tags for v0 and Devin show a trend towards AI agents that can reason about their own actions and learn from past interactions. The memory system, in particular, is a sophisticated attempt to build long-term, user-specific context.

6. Tooling Philosophy:

   • Broad & General: Devin has a vast, powerful set of general-purpose tools (shell, browser, LSP).
   • Specific & Abstracted: Replit uses highly specific tools like packager_tool and create_postgresql_database_tool that abstract away the underlying shell commands, making them more robust.
   • Parallelism: The Agent Prompt for Cursor explicitly encourages parallel tool calls, a key optimization for performance.