VibeStack

Give your AI agents the context to build not guess.
Opinionated project structure, skills, and tooling for AI-assisted development.

GitHub · Website

Install

curl -fsSL https://raw.githubusercontent.com/vibestackmd/vibestack/main/install.sh | bash

Run from the root of your project. Existing files are never overwritten.

Then run /vibestack inside Claude Code to auto-fill everything for your project.

Or install as a Claude Code plugin

If you just want the skills and hooks without project templates:

/plugin install vibestackmd/vibestack

Then run /vibestack inside any project to scaffold CLAUDE.md, Makefile, docs, and TODO.md.

What You Get

Five files:

`CLAUDE.md`	Quick-start reference card — tells any contributor (human or AI) what this project is, how it's built, how to work in it
`TODO.md`	Lightweight task tracker with a protocol for parallel work so agents don't collide
`Makefile`	Project operations via `make` — `make build`, `make test`, `make deploy TARGET=prod`. Complex scripts go in `scripts/`
`docs/`	Living knowledge base — the docs you write today prevent your AI from re-discovering the same lessons tomorrow
`.claude/skills/`	Teach AI your project's conventions and workflows via reusable skills

Skills

`/vibestack`	Analyzes your project and fills out CLAUDE.md, Makefile, docs, and TODO.md with project-specific content
`/squad`	Generates domain-specific rules and specialist subagents so Claude auto-loads the right context per file
`/todo`	Works through TODO.md tasks sequentially. `/todo populate` seeds the next batch ranked by impact
`/docs`	Captures conversation learnings into your docs folder and cleans up stale content
`/bosskey`	Scans your recent git history and generates a polished standup script you can recite to your boss — vague enough to avoid follow-ups, impressive enough to sound productive

Also includes reference skills that auto-load as context:

cli-first — teaches your AI to use platform CLIs and check .env* files instead of making raw API calls
lsp — teaches your AI to use language servers for type checking, go-to-definition, find-references, and post-change validation

Extras

CI Guards

Reusable GitHub Actions workflows: lint, test coverage, security scans, code smell checks on every PR. Supports Node/TypeScript, Python, Rust, and Go.

curl -fsSL https://raw.githubusercontent.com/vibestackmd/vibestack/main/kit/extras/ci-guards/install.sh | bash -s -- <language>

Dev Tools Installer

One-pass installer for platform CLIs (aws, vercel, etc.) and language servers (typescript-language-server, pyright, rust-analyzer, gopls) — giving your AI agent direct infrastructure access and code intelligence from the terminal.

macOS / Linux / WSL:

curl -fsSL https://raw.githubusercontent.com/vibestackmd/vibestack/main/kit/extras/dev-tools/install.sh | bash

Windows: Run the bootstrap script first to set up WSL + Ubuntu:

# PowerShell (as Administrator) Invoke-RestMethod "https://raw.githubusercontent.com/vibestackmd/vibestack/main/kit/extras/dev-tools/bootstrap-windows.ps1" | Set-Content "$env:TEMP\bootstrap-windows.ps1"; powershell -ExecutionPolicy Bypass -File "$env:TEMP\bootstrap-windows.ps1"

The `claw` Alias

The dev-tools installer offers to add claw to your shell — an alias for claude --permission-mode bypassPermissions --. Use it anywhere to run Claude uninterrupted.

claw "refactor the auth module to use JWT"

Opinions

This is what makes VibeStack different. The conventions above give your project structure — the opinions below are why.

Makefile Over Shell Scripts

Your project operations belong in a Makefile, not a bespoke shell script. make is universal, tab-completable, dependency-aware, and self-documenting. Keep targets thin — if something needs real bash logic, put it in scripts/ and call it from the target. One command to rule them all: make help.

Bypass Permissions

VibeStack ships with all Claude Code tool permissions pre-approved. No "can I run this command?" prompts. Permission prompts kill flow and add no real safety. Code quality enforcement belongs in your CI pipeline and pre-commit hooks, not in an interactive approval flow.

Squad Mode

Large codebases overwhelm AI context windows. /squad analyzes your project and breaks it into logical domains — each getting its own path-specific rules (.claude/rules/) and optionally a specialist subagent (.claude/agents/). When Claude touches a file in the auth domain, it automatically loads auth-specific conventions. Run /squad once after setup, then /squad refresh as your project grows.

CI Over Approval Gates

AI agents write bad code sometimes. The answer isn't slowing your agent down — it's a CI pipeline that catches problems automatically. Bypass mode + CI Guards = maximum speed with maximum accountability.

Finish Notification

When Claude finishes a task, your machine plays a chime and announces the project name out loud (macOS say, Linux espeak). Kick off a task, walk away, get an audio alert when it's done.

Stop vibing into chaos. Start vibing with structure. ✨

MIT License