brutal-claude-skills

18 no-mercy, adversarial-helpful skills for Claude Code.

Code reviews, roasting, design critiques, devil's advocate, pre-mortems, and more — all designed to make your work better, not your feelings warmer.

Install · Skills · Usage · Create Your Own · Contributing

Why

Claude is, by default, encouraging. That's mostly good. It's also occasionally useless when what you actually need is someone to tell you the truth about your draft, your code, your pitch, your plan, or your "Uber for X" startup idea.

This package installs skills that flip that default. Each skill puts Claude into a focused, adversarial-helpful mode for a specific task. You ask for a brutal code review, you get a brutal code review — not a pep talk.

Every skill stays on the work, not the person. Hard limits are baked in.

Install

Option 1: Run directly (no install needed)

npx brutal-review@latest init

⚠️ Always use @latest — npx caches old versions aggressively. Without @latest, you may get a stale version.

Option 2: Install globally (recommended)

npm install -g brutal-review

# Now you can just type:
brutal init
brutal help

Option 3: Local development

git clone https://github.com/premdevai/brutal-claude-skills.git
cd brutal-claude-skills
npm link

# "brutal" command now points to your local code
brutal init

Skills install to ~/.claude/skills/ by default (the directory Claude Code reads from).

⚡ Quick Usage

Once installed, use the killer feature to pipe or paste code directly into a brutal review:

# Pipe git diff straight to the reviewer
git diff | npx brutal-review use code-reviewer

# Or run interactively (paste your code)
npx brutal-review use architecture-reviewer --level 10

This will automatically launch Claude Code with the correct prompt and your input.

Command Reference

The CLI uses a clean, one-command mental model: brutal <action>. (Note: If installing globally via npm i -g brutal-review, you can just type brutal instead of npx brutal-review)

brutal init               # First-time setup (installs engineering pack)
brutal install <skill>    # Install specific skills (alias: i)
brutal use <skill>        # Interactive prompt or piped input
brutal list               # List all skills (alias: ls)
brutal status             # Check installed skills (alias: st)
brutal remove <skill>     # Remove skills (alias: rm)
brutal doctor             # Run diagnostics
brutal export cursor      # Export to Cursor (.cursorrules)
brutal hook git --all     # Install all 3 git hooks
brutal hook github        # Add GitHub Actions PR review
brutal upgrade            # Check for updates

Skill Packs

Don't want everything? Install curated bundles:

brutal install pack <name>

Team Install (Project Level)

Install skills at the project level so your whole team gets them via git:

# Install to .claude/skills/ in your current project
brutal install all --project

# Commit to share with your team
git add .claude/skills/ && git commit -m "chore: add brutal claude skills for the team"

Check what's installed globally vs. in the current project:

brutal status

The Skills

— review your work product

— interrogate your decisions

— survive the real thing

— pure chaos

Usage

Once installed, you don't need to do anything special. Each skill has a description that tells Claude when to activate. Just ask naturally:

"Review my code"              → brutal-code-reviewer
"Roast my Twitter bio"        → roast-mode
"Play devil's advocate"       → devils-advocate
"What could go wrong?"        → pre-mortem
"Is this BS?"                 → bs-detector
"What am I assuming here?"    → assumption-auditor
"Review my README"            → brutal-readme-reviewer

You can also name a skill explicitly if you want a specific one.

Brutality Scale (0–10)

Every skill supports a brutality level from 0 to 10. Default is 7. Just tell Claude how hard to go:

"Review my code, level 3"           → professional, no sarcasm
"Review my code"                    → default (7), savage
"Review my code, go nuclear"        → level 10, profanity unlocked
"Be gentle with this one"           → level 1–2
"Maximum brutality"                 → level 10
"Turn it down"                      → drops 3 levels mid-conversation

🔗 Integrations — Cover the Entire Dev Cycle

Brutal can embed itself into every stage of your workflow — from keystroke to merged PR.

Git Hooks (local, per-repo)

Install in one command from inside any git repo:

# Install all 3 hooks at once
npx brutal-review hook git --all

# Or pick specific hooks
npx brutal-review hook git --commit-msg   # Block lazy commit messages
npx brutal-review hook git --pre-commit   # Scan staged diff for secrets & smells
npx brutal-review hook git --pre-push     # Guard against pushing to main

# Uninstall (non-destructive — restores your originals)
npx brutal-review hook uninstall

What each hook does:

To skip a hook for a single commit (use carefully):

git commit --no-verify

GitHub Actions (automated PR review)

npx brutal-review hook github             # Default level 8
npx brutal-review hook github --level 10  # Nuclear

This generates .github/workflows/brutal-review.yml. Every PR gets a brutal code review posted as a comment. Old reviews are automatically replaced — no spam.

git add .github/workflows/brutal-review.yml
git commit -m "ci: add brutal code review workflow"
git push

[!IMPORTANT]

Export to Cursor & Copilot

You can use these brutal personas in other AI coding assistants. Export them as a .cursorrules file for Cursor, or a .github/copilot-instructions.md file for GitHub Copilot.

# Export the engineering pack to Cursor
npx brutal-review export cursor --pack engineering

# Export a specific skill to GitHub Copilot
npx brutal-review export copilot brutal-code-reviewer

Programmatic API

Building a custom PR review bot? An internal Slack integration? You can consume the prompts programmatically via Node.js:

const { getSkill, listSkills } = require('brutal-review');

// Get all available skill names
const skills = listSkills();

// Get the raw prompt for a specific skill, overriding the default brutality level
const prompt = getSkill('brutal-code-reviewer', { 
  stripFrontmatter: true, 
  level: 10 
});

console.log(prompt); // Use this in your Anthropic API calls

Manual Install (no npm)

# Clone the repo
git clone https://github.com/premdevai/brutal-claude-skills.git
cd brutal-claude-skills

# Copy the skills you want
cp -r skills/brutal-code-reviewer ~/.claude/skills/
cp -r skills/roast-mode ~/.claude/skills/

# Or copy everything
cp -r skills/* ~/.claude/skills/

Uninstall

# Remove a specific skill
npx brutal-review remove brutal-code-reviewer

# Remove all installed skills
npx brutal-review remove all

# Remove git hooks
npx brutal-review hook uninstall

Creating Your Own Skill

Each skill is one folder with one SKILL.md file:

my-skill/
└── SKILL.md

The SKILL.md uses YAML frontmatter followed by the skill body in markdown:

---
name: my-skill
description: A one-line description. This is what Claude reads to decide when
  to trigger the skill, so include concrete trigger phrases.
---

# My Skill

## Persona
Who the skill embodies...

## Behavior Rules
What it always/never does...

## Output Format
How it structures responses...

See any skill in skills/ for a real example, or read the full Skill Authoring Guide.

Design Principles

1. Stay on the work, never the person. Roasts target choices, not identity.
2. Quote what you attack. Every critique cites the actual line, bullet, or sentence.
3. Cut before rewriting. Most things are too long. Most fixes start with deletion.
4. Hard limits are non-negotiable. No punching at race, gender, sexuality, disability, mental health, body, or trauma — even in roast-mode.
5. No participation trophies. If something's bad, it's bad.

Contributing

Contributions are welcome! Whether it's a new skill, a bug fix, or an improvement to an existing skill.

See CONTRIBUTING.md for guidelines.

License

MIT — do whatever you want with it.

A Note on Tone

If a skill ever crosses a line — punches at identity instead of choices, hands out cruelty without a joke, or refuses to drop the bit when you tap out — that's a bug, not a feature. Open an issue.

The goal is to make your work better. Not to be liked.