Chatmode Fine-tuning: Spec-driven Copilot Modes

Pragmatic system prompts for AI agents that value clarity over hype.

Table of Contents

• Overview
• What's Inside
  • chatmode-v1.0
  • chatmode-v2.0
  • chatmode-v3.0
• Why This Exists
• How to Use These Prompts
• Design Philosophy
• Contributing
• License

Overview

This repository contains a collection of chatmode system prompts designed to guide GitHub Copilot-style agents toward spec-driven, stateful, and pragmatic behavior when working on real software projects.

These prompts intentionally avoid flashy buzzwords and emoji-heavy "vibe coding" instructions. Instead, they focus on established software engineering practices and reproducible workflows that work in professional team environments.

What's Inside

chatmode-v1.0

Version 1.0 is the initial implementation focusing on making specific LLM models spec-driven rather than using GitHub Copilot's default Agent mode behavior.

File	Target Model
aiconf-gpt-4.1.md	OpenAI GPT-4.1
aiconf-gpt-5-mini.md	OpenAI GPT-5-Mini
aiconf-grok-code-fast.md	Grok Code Fast
aiconf-sonnet.md	Claude Sonnet

All prompts in v1.0 share the same foundation:

• State synchronization at the start of every session
• Spec-driven development (plan before coding)
• Collaborative execution with explicit task tracking
• Scalability strategies for handling large codebases

chatmode-v2.0

Version 2.0 builds on v1.0 with significant improvements for autonomous, tool-enabled workflows in VS Code.

File	Target Model
aiconf-gpt-4.1.md	OpenAI GPT-4.1
aiconf-gpt-5-mini.md	OpenAI GPT-5-Mini
aiconf-grok-code-fast.md	Grok Code Fast
aiconf-sonnet.md	Claude Sonnet

Key improvements in v2.0:

• Full Tool Support: Expanded tools list to include all VS Code capabilities (terminal selection, test files, search results, codebase navigation, etc.)
• Autonomous Execution: Enhanced agent behavior to work continuously until problems are fully solved, only yielding when complete
• Internet Research: Integrated web fetching and recursive link gathering for up-to-date documentation and dependency research
• Enhanced Verification: Automatic lint checks, test execution, and error reporting after code changes
• Resume Capability: Explicit support for "resume" or "continue" commands to pick up from the last incomplete todo item
• Tool-First Approach: Stronger emphasis on using VS Code tools for all operations (editing, testing, debugging) rather than manual steps
• Iterative Problem-Solving: Built-in workflow for testing edge cases, debugging root causes, and validating solutions rigorously

When to use v2.0: If you need autonomous problem-solving with extensive research, rigorous testing, and minimal manual intervention, use v2.0. If you prefer more frequent check-ins and explicit approval at each step, stick with v1.0.

chatmode-v3.0

Version 3.0 represents a complete rewrite optimized for specific AI model families, incorporating official prompting best practices from OpenAI and Anthropic.

File	Target Model Family
gpt-finetuned.md	OpenAI GPT models (gpt-4.1, gpt-5)
sonnet-finetuned.md	Anthropic Claude/Sonnet models

Version 3.0 introduces model-specific optimizations:

For OpenAI GPT Models (gpt-finetuned.md)

• OpenAI Prompting Guide Alignment: Follows OpenAI's GPT-5 prompting recommendations including tool preambles, agentic persistence, and context gathering heuristics
• Conversation Summaries: Replaces session_state with dynamic conversation summaries at task start
• Context Gathering Optimization: Parallel search strategies with early stop criteria to balance thoroughness and efficiency
• Tool Preambles: Clear communication before, during, and after tool usage for better user experience
• Comprehensive Tool Support: 40+ VS Code built-in tools including #fetch, #githubRepo, #codebase, #problems, and all editing/testing tools
• Model compatibility: gpt-finetuned.md has been tested with the raptor-mini model in GitHub Copilot and validated to work as expected; use raptor-mini for fast, efficient editing and tool-enabled workflows.

For Anthropic Claude/Sonnet Models (sonnet-finetuned.md)

• XML-Structured Thinking: Extensive use of XML tags (<thinking>, <analysis>, <plan>, <answer>) for structured reasoning
• Chain of Thought (CoT): Explicit instructions for step-by-step reasoning with visible thought processes
• Role Prompting: Clear role definition as seasoned software engineer and architect
• Prefilling Support: Guidance on using Anthropic's prefilling feature to skip preambles and control output format
• Hierarchical XML Tags: Nested XML structure for all major sections following Anthropic's recommended patterns

Common v3.0 Features

• Beast Mode Philosophy: Autonomous, action-oriented tone inspired by proven chatmode patterns
• Mandatory Web Research: Heavy emphasis on using #fetch tool for up-to-date information before any code changes
• Deprecation Awareness: Explicit instructions to avoid deprecated APIs (e.g., Go's ioutil) and verify current best practices
• Frontend Defaults: Tailwind CSS mandatory for all frontend work unless specified otherwise
• Rigorous Testing: Multiple test iterations to catch edge cases before considering tasks complete
• No Session State Files: Both models use conversation summaries instead of persistent .session_state files
• Tool-First Development: Prefer VS Code tools over manual steps for all operations

When to use v3.0: Use v3.0 when working with modern GPT or Claude models and you want cutting-edge, model-specific optimizations based on official vendor recommendations. v3.0 is recommended for all new projects.

Note

v3.0 files are named differently (gpt-finetuned.md and sonnet-finetuned.md) to emphasize they are optimized for specific model families rather than individual model versions.

Why This Exists

Most publicly shared chatmodes lean into dramatic language, excessive emojis, and hyperbolic claims about "superhuman" or "quantum cognitive" workflows. From a developer's perspective—especially those who learned software engineering before the current AI agent craze—that approach can feel gimmicky and unhelpful.

Real projects need real processes:

• Clear specifications
• Versioned implementation plans
• Verifiable, reviewable changes
• Proper handover documentation

This repo brings AI agents back to fundamentals: spec-driven development, stateful sessions, and practical collaboration.

A Note on "Vibe Coding"

The problem emerges when people create chatmodes without understanding software engineering fundamentals, leading to:

• Unmaintainable spaghetti code
• Confusing project histories
• Poor team collaboration
• Projects that can't scale beyond prototypes

If you're building chatmodes for professional or collaborative projects, understanding Software Development Life Cycle (SDLC) is essential.

Tip

Recommended reading

Software Engineering: A Practitioner's Approach by Roger S. Pressman

How to Use These Prompts

1. Choose the version that matches your workflow:

   • v1.0 for collaborative, step-by-step approval at each task
   • v2.0 for autonomous, tool-enabled problem-solving with minimal intervention
   • v3.0 for model-specific optimizations with official vendor best practices (recommended)

2. Select the prompt file that corresponds to your AI model:

   • v3.0 (recommended):
     • Using OpenAI GPT-4.1 or GPT-5? → chatmode-v3.0/gpt-finetuned.md
     • Using Anthropic Claude Sonnet? → chatmode-v3.0/sonnet-finetuned.md
   • v2.0:
     • chatmode-v2.0/aiconf-gpt-4.1.md, aiconf-gpt-5-mini.md, aiconf-grok-code-fast.md, or aiconf-sonnet.md
   • v1.0:
     • chatmode-v1.0/aiconf-gpt-4.1.md, aiconf-gpt-5-mini.md, aiconf-grok-code-fast.md, or aiconf-sonnet.md

3. Load it as a custom instruction or system prompt in your GitHub Copilot chatmode configuration or custom chat mode settings

4. Start your session:

   • v3.0: Simply start working—the model will synthesize conversation summaries automatically
   • v1.0/v2.0: Provide a session_state summary containing high-level goal, approved specification, implementation plan, and current progress

5. Follow the workflow outlined in the prompt for consistent, spec-driven development

Design Philosophy

This project follows these core principles:

1. Clarity over hype — Professional, clear instructions without unnecessary drama
2. Spec-driven development — Always plan before implementing
3. Tool-first approach — Leverage VS Code tools for all operations
4. Research-enabled — Use web research to stay current with dependencies and best practices
5. Model-specific optimization — v3.0 tailors prompts to each AI vendor's recommended patterns
6. Reproducible workflows — Processes that work in team environments
7. Rigorous verification — Test thoroughly before considering work complete

Contributing

Contributions are welcome. Please open issues or pull requests with clear descriptions of proposed changes.

When contributing:

• Maintain the professional, no-nonsense tone
• Include rationale for prompt changes
• Test changes with actual coding sessions before submitting
• Document which model versions you tested with

License

These files are provided as-is, with no warranty. Use them however you'd like, responsibly.