Agentic Dev Starter

A minimal framework for effective AI-assisted development.

TLDR

A simple framework to maintain context between AI assistant sessions using just two files (DEVSTATE.md, PLAN.md) and a prompt (PROMPT.md). Helps keep AI-assisted development focused and efficient by preserving state between sessions while keeping each interaction targeted on specific tasks and goals.

Contents

• Purpose
• The Problem
• A Solution
• The Files
• Getting Started
• Requirements
• Development Continuity Prompt
• Usage Example
• Contrib

Purpose

This micro-framework provides a simple set of files and a prompt template to streamline LLM-augmented development workflows.

(Tested with Cline and Claude 3.x Sonnet/Haiku)

The problem

When working with chat-based coding agents, the length of the prompt increases with every message in the same session, making each request progressively more expensive.

Long-running chats can reach a point of "diminished returns" when focus, direction, or subject changes within the chat.

It makes sense to work with short sessions focused on atomic tasks. However, when working toward a larger goal, we need to retain state between tasks by:

• Prompting the agent on how to start and finish tasks
• Prompting the agent on how to read and write state

A solution

This framework uses just two files (DEVSTATE.md, PLAN.md), one tool (git), and one prompt (in PROMPT.md) that you add to your agent's settings to be read on every request.

The files and their purpose

• DEVSTATE.md: The LLM overwrites this file after completing each task to record the current development status for the next session.
• PLAN.md: User-managed file detailing short-term goals, constraints, requirements, and focused tasks. The LLM may update this file, but the user ensures the content remains concise and focused.
• PROMPT.md: Contains the prompt to add to your AI assistant's settings.

Getting Started

1. Choose your approach:

   # Either clone this starter repository
   git clone /~https://github.com/justinphilpott/agentic-dev-starter.git
   cd agentic-dev-starter
   
   # Or copy the files to your existing project
   # Just copy PLAN.md and DEVSTATE.md to your project

2. Set up your project:

   • Customize PLAN.md with your project goals, constraints, and initial tasks
   • Initialize DEVSTATE.md with your project's starting state

3. Configure your AI assistant:

   • Copy the prompt from PROMPT.md into your AI assistant's settings
   • Ensure your assistant has access to read and write files in your project

Requirements

• Git
• VSCode or similar dev environment that supports chat-based coding agents
• An agentic coding assistant that allows custom prompts in settings

Usage Example

1. Initialize your project:

   • Create PLAN.md with your initial goals and tasks
   • Create an empty DEVSTATE.md file

2. Start your first development session:

   • Open your project in VSCode with your AI assistant
   • Ask your assistant to help with a specific task from PLAN.md
   • The assistant will follow the Task Start Protocol to understand the current state

3. Complete the task:

   • Work with the assistant to implement the solution
   • The assistant will follow the Task Completion Protocol
   • DEVSTATE.md will be updated with the current status
   • A git commit will be suggested

4. Continue development:

   • In your next session, the assistant will read DEVSTATE.md to understand what's been done
   • Select the next task from PLAN.md
   • Repeat the process

This approach maintains context between development sessions while keeping each interaction focused and efficient.

Contrib

Feedback, ideas, and pull requests are welcome.