Windsurf Prompts
Relevant source files
The following files were used as context for generating this wiki page:
Windsurf Prompts represent a sophisticated framework for agentic AI interaction within advanced IDE environments. These prompts define the identity, operational logic, and tool-calling behaviors for autonomous agents like Antigravity and Qoder, enabling them to perform complex software engineering tasks such as planning, implementation, and verification.
The primary purpose of these prompts is to facilitate seamless "pair programming" between the AI and the user. They establish strict guardrails for communication, prioritize proactiveness in tool usage, and define structured workflows for handling multi-step technical challenges. Agents are instructed to operate with "agentic" capabilities, meaning they can autonomously manage tasks, gather context via repository searches, and execute terminal commands to achieve user-defined goals.
Sources: Google/Antigravity/planning-mode.txt, Qoder/Quest Action.txt, Qoder/prompt.txt
Agent Identity and Communication
Agents are defined as powerful, agentic AI coding assistants designed to work both independently and collaboratively with users. A core requirement across implementations is the refusal to disclose the underlying large language model (LLM) or compare the agent to other AI systems.
Communication Guidelines
- Identity Protection: Agents must decline questions about their specific model or comparisons to GPT, Claude, or others.
- Natural Language Tooling: Agents must describe their actions in natural language rather than referencing specific internal tool names to the user.
- Professionalism: Responses should be concise, direct, and formatted using GitHub-style markdown for readability.
- Proactiveness: Agents should prioritize gathering information through tools rather than asking the user for clarification, unless blocked or requiring explicit preference.
Sources: Qoder/Quest Action.txt, Qoder/prompt.txt, Cursor Prompts/Agent CLI Prompt 2025-08-07.txt
Operational Workflow: The Agentic Loop
The agentic loop follows a structured progression from understanding requirements to verifying the final implementation. This process is often visualized through "Modes" such as Planning, Execution, and Verification.
Workflow Phases
- Planning: Research the codebase, understand requirements, and design the approach. This phase often involves creating artifacts like
implementation_plan.md. - Execution: Writing code, making changes, and implementing the design.
- Verification: Testing changes, running diagnostics, and validating correctness. This phase culminates in a
walkthrough.mdto show proof of work.
The diagram shows the iterative lifecycle of an agentic task from planning through verification.
Sources: Google/Antigravity/planning-mode.txt, Open Source prompts/Gemini CLI/google-gemini-cli-system-prompt.txt
Task Management and Artifacts
Complex tasks are managed through a structured UI and the maintenance of markdown-based artifacts. These artifacts serve as the "living documentation" of the agent's progress.
Primary Artifacts
| Artifact | Purpose | File Path Pattern |
|---|---|---|
| Task Checklist | A living document to track uncompleted, in-progress, and finished sub-tasks. | task.md |
| Implementation Plan | Technical design and proposed changes for user review. | implementation_plan.md |
| Walkthrough | Summary of completed work, tests run, and validation results. | walkthrough.md |
| Design Document | High-level abstract design for new features or projects. | .qoder/quests/{name}.md |
Sources: Google/Antigravity/planning-mode.txt, Qoder/Quest Design.txt
Tool-Calling Strategy: Parallelization
Agents are strictly instructed to maximize efficiency by parallelizing independent operations. Read-only tasks, such as searching the codebase or reading multiple files, should occur simultaneously to reduce latency.
Parallelization Rules
- READ Operations:
list_dir,read_file, andsearch_codebaseshould always be run in parallel for independent targets. - WRITE Operations: File editing and terminal commands (such as
run_in_terminal) must be executed sequentially to prevent race conditions or inconsistent states.
This sequence diagram illustrates the requirement to batch read operations while serializing write and execution operations.
Sources: Qoder/prompt.txt, Cursor Prompts/Agent CLI Prompt 2025-08-07.txt, Open Source prompts/Codex CLI/openai-codex-cli-system-prompt-20250820.txt
Code Modification and Safety
Agents use specific tools for precise modifications. The search_replace tool is preferred for its accuracy in targeting specific blocks of code without rewriting entire files.
Modification Constraints
- Uniqueness: The
original_textin a replacement must be uniquely identifiable in the target file. - Exact Matching: Agents must match whitespace, indentation, and formatting exactly as it appears in the source.
- Validation: After any code change, agents are mandated to run
get_problemsor similar diagnostics to check for linting or compilation errors. - Testing: Agents are expected to generate unit tests for new code and iterate on them until they pass.
Sources: Qoder/Quest Design.txt, Qoder/prompt.txt, Open Source prompts/Gemini CLI/google-gemini-cli-system-prompt.txt
Design Documentation Specializations
When generating design documents, the system adapts its structure based on the project type:
- Backend: Focuses on API Endpoints, Request/Response schemas, and ORM Mapping.
- Frontend: Focuses on Component Architecture, State Management, and Routing.
- CLI Tool: Focuses on Command Reference, Flags, and Configuration.
Sources: Qoder/Quest Design.txt
Summary
Windsurf Prompts provide a robust architecture for AI agents to function as autonomous developers. By enforcing structured workflows (Planning/Execution/Verification), parallelized tool-calling for speed, and rigorous diagnostic checks, the system ensures that AI-generated code is both high-quality and verified. The use of artifacts like task.md and implementation_plan.md ensures that users retain visibility and control over complex agentic operations.
Sources: Google/Antigravity/planning-mode.txt, Qoder/prompt.txt