GitHub - ulrischa/agentic-coding-biolerplate: A folder structure and files to start an ai assisted agentic coding project

GitHub

3 min read Original article ↗

Agentic Coding Boilerplate

A lightweight, language-neutral project starter for GitHub Copilot in VS Code, OpenAI Codex and Claude Code. It provides shared project instructions, reusable review workflows and documentation templates without imposing a framework or adding runtime dependencies.

Contents

File / directory Purpose
AGENTS.md Shared working conventions, safety rules and verification expectations
CLAUDE.md Imports the shared instructions for Claude Code
.agents/skills/ Canonical code-review and security-review workflows
.claude/skills/ Claude Code entrypoints pointing to the canonical workflows
.editorconfig Shared text formatting conventions
.gitignore Common exclusions for local configuration and generated files
.vscode/extensions.json GitHub Copilot extension recommendations
templates/README.md README skeleton for the project you build from this starter
docs/ARCHITECTURE.md Architecture documentation skeleton
docs/DECISIONS.md Lightweight template for significant technical decisions
docs/BOILERPLATE.md Detailed setup and usage guide
src/, tests/ Optional source and test directory placeholders

Getting started

  1. Clone or download this repository, or use GitHub's Use this template option if available. When adding it to an existing project, reconcile existing files and instructions before copying anything over them.
  2. Open the repository root in VS Code and use your chosen coding assistant. The starter itself requires no WSL, administrator privileges or package install. Install and authenticate the assistant separately; its access requirements apply.
  3. Replace this README with the contents of templates/README.md. Fill in your project purpose, supported runtime versions, setup instructions and real verification commands. The separate template keeps this repository's documentation out of your application's README.
  4. Adapt AGENTS.md, .gitignore and .editorconfig to the project. Complete architecture and decision records only where useful; adjust or remove the placeholder directories as needed.
  5. Start a new agent session at the repository root. Check that the shared instructions and review skills are available before assigning a small task.

The guide includes a copyable setup request and a completion checklist, and explains file responsibilities, assistant discovery, troubleshooting and how to add a skill. Keep it as a reference while replacing this README with your own project documentation.

Using the review skills

Workflow GitHub Copilot Codex Claude Code
Code review Ask to use project-code-review $project-code-review /project-code-review
Security review Ask to use project-security-review $project-security-review /project-security-review

Specify the scope, such as uncommitted changes or the last commit. Both workflows report findings and suggested fixes; applying fixes requires a request that covers that work. Automatic skill selection depends on the assistant and is not guaranteed.

The project- prefix avoids collisions with bundled commands. Claude entrypoints reuse the canonical files without symlinks, keeping ordinary Windows checkouts simple. Copilot may discover both entrypoints; they lead to the same workflow.

Scope and verification

This repository contains configuration and documentation, not an application. Application test, lint, build and run commands are not configured. Add commands appropriate to your project rather than treating placeholders as successful checks.

Local file links, skill metadata and entrypoint paths, JSON syntax, ignore patterns and text formatting were checked. Full interactive sessions across all three assistants were not executed. Discovery and optional features depend on the installed versions and settings; see the official references in the guide.

Custom agents, hooks, MCP servers, CI pipelines and additional prompt files are optional extensions for concrete needs. Markdown instructions and ignore patterns are not access controls; actual access depends on tool permissions and the environment.