GitHub - daitangio/take-ai-control: Take AI Control back template repository.

This repository is a template to govern your AI chats with more grip than usual. We propose a tailored AGENTS.md and an optional ready-to-use container for proper isolation. A set of demo/* branches show real application developed with AI.

Project leverages PI.dev coding agent and an aggressive isolation done with Docker + VS Code

Tested on Linux and MacOS

Folder Structure:

.
|-- .devcontainer            <- Visual studio code dev container
|-- bin/runInContainer.sh    <- Zero-lock-in launcher for running outside vscode
├── .agents/skill            <- Basic Skills with compatibility between codex, claude and copilot
├── README.md    
├── etc                      <- scripts used by Docker to setup your environment
└── var                      <- Contains configuration not meant to be versioned
    ├── pi-agent             <- Contains pi.dev configuration
    │   └── models.json      <- Models to be configured
    └── pi-sessions          <- Contains pi.dev session

Basic principles:

• Liquid Models: ability to switch between models and model providers. Easier to do with pi.dev. See also liquid-models.md
• A set of basic skills
• Isolated Dev container for stronger security
• Experimental SDLC feature/* branches (see below)

Liquid Models

• Agents skills can be used on every major harnesses (codex, claude, copilot, pi.dev)
• pi.dev installation retain sessions and config inside the var directory
• Tested on copilot, codex and cloude. Currently Claude code is only tested with DeepSeek integration (see deep seek manual, use env variable in your devcontainer.env)

SDLC

AI development increase so much the speed you need to change the way you develop. In particular you need some directive to

• avoid the human to be the bottleneck

• something to track modification and work in teams. There are a lot of tools, and we are testing them.

• demo/nello-openspec

  A trello clone based on OpenSpec OpenSpec Pro: collect specification, light and fast and create a Soure-of-truth spedification. Contra: (none found yet)

• feature/ai-sdlc-{copilot,pi,claude}

  Based on https://github.com/awslabs/aidlc-workflows#github-copilot These variants provide true simple AI-SDLC workflow without the need of Kiro Pro: very clear in proposition and pros/cons Contra: very slow, do not seems very effective if you do not have deep pocket (token cost is a high).

Getting started

• Clone or fork this project and use as template
• Define a .devcontainer/devcontainer.env with all your API keys (i.e. DeepSeek, Claude etc) if you have already This file is common to all the containerized approaches
• Three options:
  • run Visual Studio Code DevContainer mode. If so:
    • Review .devcontainer/devcontainer.json
  • run ./bin/runInContainer.sh to get a throwaway container on the command line If you do not have claude code installed, use this method to ensure some empty folder are created
  • Use without container (but please avoid pi.dev in this scenario)

Using pi.dev

Pi.dev is fantastic because it never will ask for command confirmation: but it is also a risk.

Once you have your terminal, install your Pi.dev's preferred extensions (you need to do this just once): For instance try

    pi install git:github.com/jonjonrankin/pi-caveman
    pi install npm:pi-subagents

The extensions will be stored in the var/pi-agent subdirectory (see above). After that, you can use pi.dev as you wish. Our suggestion is to get accustomed to pi.dev with a short session, then you can look forward on the chapter included in the doc directory like SUBAGENTS (this part is a Work in progress WIP).

Provided Skills

A compact and curated list of skills are provided. Main goal is to readuce tokens:

• AGENTS.md provide usage of rtk tool. CLAUDE.md is symlinked to it as good practice to share it
• Under .agents (symlinked to .claude for ClaudeCode) a mark-it-down converter is provided, to convert documents in a more compact and mangeable form.