AI-Code-Framework

This repository is a framework of reusable project bases for working with Claude and other coding agents without repeating a giant master prompt on every task.

The four bases

• Dev-Project - generic project bootstrap and resume flow for software projects.
• PowerShell-Module - production-grade PowerShell module engineering.
• Containers - container-first projects, Docker/Compose/Swarm guidance, and portable CI specification.
• Puppet - production-grade Puppet module engineering with secure-by-default controls and reusable scaffolding.

The bases solve different problems:

• Dev-Project is the generic bootstrap and resume base for software projects.
• PowerShell-Module is the production-grade base for PowerShell module engineering.
• Containers is the container-first base for Docker, Compose, Swarm, and portable CI.
• Puppet is the production-grade base for Puppet module architecture, data modeling, testing, and release governance.

None of these bases is a lower or higher tier in value. They are different tools for different project goals.

Each base follows the same pattern:

• a thin entry prompt for new work
• a thin resume prompt for existing work
• one canonical skill that holds the operating contract
• topic-focused references loaded only when relevant
• ready-to-copy templates for the actual project scaffold
• optional CI specification plus CI generator pair when the domain benefits from it

How to use it

• Start a new project from the base that matches the problem domain.
• Resume an existing project from its resume prompt and live docs.
• Keep the context window small and work in small iterations.

Why this structure works

• The prompt stays short and cheap.
• The rules stay in one canonical skill instead of being duplicated everywhere.
• The templates stay practical and executable instead of being abstract.
• The same pattern works across generic apps, PowerShell modules, and container projects.
• The framework can stay generic at the root while each base adds domain-specific structure only where it pays off.
• The token-saving model is deliberate: keep the context window small, use /compact when the thread gets dense, and restart with a new chat when the conversation is no longer efficient.

Skills

Think of the skills/* folders as reusable skill templates. For Claude Code to auto-discover them, the folder must be copied into a real skill location such as:

• ~/.claude/skills/<skill-name>/SKILL.md for personal use
• .claude/skills/<skill-name>/SKILL.md for repo-scoped use

Copied there, Claude loads the SKILL.md frontmatter and the skill becomes active. Kept in this repository, it remains a reusable source template.

Cross-agent model

Each generated project should use one canonical instructions file plus thin pointers for each agent.

Default pattern used by all bases:

• AGENT.md - canonical source of truth
• CLAUDE.md - Claude pointer to AGENT.md
• CODEX.md - Codex pointer to AGENT.md
• DEEPSEEK.md - DeepSeek pointer to AGENT.md
• .github/copilot-instructions.md - Copilot pointer to AGENT.md

This pattern keeps the rules single-sourced and easy to resume across conversations and tools.

Default enterprise policy in this framework:

• Deliverables language: English only.
• Execution mode: full audit before fixes, with explicit "Not Verified" reporting when a check cannot run.