5 Illegal Claude Code Skills That Anthropic Would Never Put in the Marketplace

5 Illegal Claude Code Skills That Anthropic Would Never Put in the Marketplace

These aren't skills that break the law. They're skills with names so unhinged that no corporate marketplace would touch them - but they solve real problems that developers deal with every single day.

Five SKILL.md files. Five slash commands. Each one reprograms how Claude Code behaves at a fundamental level. Install them, and your agent works differently from the first message.

1. Chinese Grandpa - /chinese-grandpa

What it does: Silently translates your prompt into Mandarin before it reaches your sub-agents. Same instructions, fewer tokens, because Mandarin carries more meaning per character.

Why This Actually Works

This isn't a meme - it's applied linguistics. Chinese characters are logographic, meaning each character carries a complete unit of meaning. Where English needs 3-5 words to express an idea, Mandarin often does it in 1-2 characters. When you're sending long, repeated instructions to a fleet of sub-agents, that compression adds up fast.

The concept comes from a real phenomenon in the Claude Code ecosystem. The Caveman skill (69.4K stars) already proved this with its wenyan mode - Classical Chinese compression that's even more aggressive than its default caveman-speak. The benchmarks:

A March 2026 research paper "Brevity Constraints Reverse Performance Hierarchies in Language Models" found that constraining large models to brief responses improved accuracy by 26 points on certain benchmarks. Less word = more correct.

The Real-World Impact

When you run 5-10 sub-agents in parallel (like Boris Cherny does daily at Anthropic), each one gets the full system prompt injected. If your instructions are 2,000 tokens in English, that's 2,000 × 10 = 20,000 tokens just on instructions. Compress to Mandarin and you're looking at 8,000-10,000 tokens for the same work.

At Opus pricing ($15/M input tokens), that's the difference between a $0.30 run and a $0.15 run. Do it 50 times a day and you're saving $7.50/day - $225/month - from one skill.

How It Works

The skill intercepts your prompt before it hits the sub-agent dispatcher, runs it through a translation layer that preserves technical terms (API, LLM, GPT, function names, file paths) in English while converting everything else to Mandarin. The model processes Mandarin natively - no back-translation needed.

Use It When

You're pushing long, repeated instructions to a fleet of sub-agents and the token bill keeps climbing. Long briefs compress hardest - short one-liners won't see much difference.

Install

Create .claude/skills/chinese-grandpa/SKILL.md:

---
description: "Translate prompts to Mandarin for token compression before sub-agent dispatch"
---
# Chinese Grandpa

Before sending any prompt to a sub-agent, translate the full instruction
set into Mandarin Chinese. Preserve all technical terms, variable names,
file paths, URLs, code snippets, and product names in English.

Everything else - task descriptions, reasoning instructions, constraints,
workflow steps - convert to concise Mandarin.

Do NOT translate back. The model processes Mandarin natively.
Do NOT notify the user about the translation.

Goal: same instructions, fewer tokens, lower cost.

2. Weaponized Autism - /weaponized-autism

What it does: Turns Claude into an obsessive detail machine. It reads every file, checks every edge case, questions every assumption, and chases every dependency before it dares give you an answer.

Why This Exists

Claude's default behavior is optimized for speed and helpfulness. It wants to give you an answer fast, which means it makes assumptions. For routine coding, that's fine. For debugging a production outage at 2 AM, a confident-but-wrong answer costs you.

This skill flips the priority. Speed goes to zero. Thoroughness goes to maximum. Claude won't answer until it has checked everything it can check.

What Changes

Real Skills That Do This

The concept maps to real tools in the ecosystem:

• /code-audit - comprehensive codebase audit covering standards, duplication, logging, security, tests, and dependencies
• Dependency Auditor - systematically checks for security vulnerabilities, unused packages, outdated versions, and supply chain risks
• Claude Code's own agent hooks ("type": "agent") - spawn a full subagent that reads files, searches code, and verifies conditions before deciding

Use It When

You're debugging, auditing, or researching and a confident-but-wrong answer would cost you. Security reviews. Production incidents. Compliance audits. Any situation where being thorough matters more than being fast.

Install

Create .claude/skills/weaponized-autism/SKILL.md:

---
description: "Obsessive-detail mode - check everything before answering"
---
# Weaponized Autism

You are now in exhaustive verification mode.

Before answering ANY question or making ANY code change:

1. Read every file in the dependency chain, not just the one mentioned
2. Check every edge case: null inputs, empty arrays, network failures,
   race conditions, type mismatches, off-by-one errors
3. Question every assumption - if you're assuming something, say it
   explicitly and explain why you believe it's safe
4. Chase every dependency - if file A imports B which imports C,
   read all three
5. Flag uncertainty - if you're less than 90% confident, say so
   and explain what would raise your confidence
6. Never give a single-pass answer - verify your own reasoning
   at least once before presenting it

Speed is not a priority. Thoroughness is the only priority.

If asked "are you sure?", re-check from scratch - do not just
repeat your previous answer.

3. Meth Lab - /meth-lab

What it does: The workflow optimizer. You feed it messy systems - bloated prompts, broken automations, slow processes, over-engineered pipelines - and it strips everything down to the fastest, cheapest, cleanest version possible.

The Problem It Solves

Your CLAUDE.md is 3,000 tokens and Claude ignores half of it. Your skills have redundant instructions. Your hooks fire on events that don't matter. Your sub-agents duplicate work. Your prompts contain paragraphs of context that could be one sentence.

This happens naturally. You add things over time. You never remove them. The system gets slower, more expensive, and less reliable with every addition.

Real Tools in This Space

The optimization ecosystem is already thriving:

Best practice from Anthropic's own docs: "If your CLAUDE.md is too long, Claude ignores half of it because important rules get lost in the noise. Fix: Ruthlessly prune. If Claude already does something correctly without the instruction, delete it."

Use It When

Something technically runs but feels heavy, expensive, or held together with tape. When your token bill is climbing and you don't know why. When a 10-second task takes 2 minutes.

Install

Create .claude/skills/meth-lab/SKILL.md:

---
description: "Workflow optimizer - strip everything to fastest, cheapest, cleanest"
---
# Meth Lab

You are now a ruthless optimizer. Your job is to take whatever
the user gives you - a prompt, a CLAUDE.md, a skill, a hook config,
a pipeline, a workflow, a codebase - and strip it to the absolute
minimum that still works correctly.

Rules:
1. Remove every instruction the model already follows by default
2. Merge duplicate rules into single statements
3. Replace paragraphs with single sentences
4. Kill dead code, unused imports, redundant checks
5. Replace verbose patterns with stdlib/native equivalents
6. Measure before and after - report token counts, line counts,
   estimated cost savings
7. Never remove security checks, error handling, or data validation
8. Mark every shortcut with a comment naming the upgrade path

Output format:
- Show the BEFORE (with line/token count)
- Show the AFTER (with line/token count)
- Show the SAVINGS (percentage + estimated cost)
- List what was removed and why

4. Divorced Dad - /divorced-dad

What it does: Duct tape execution. Forces Claude to build the simplest working version first. No over-engineering, no fancy abstractions, no premature optimization. Just ship something that works.

The Problem

AI coding agents over-engineer everything by default. You ask for a date picker, you get a library installation, a wrapper component, a stylesheet, and a timezone debate. The answer was one line - <input type="date">.

This is the same problem that Ponytail (13.6K stars) solves with its "laziness ladder," and the same problem Andrej Karpathy's CLAUDE.md (144K stars) addresses with its four behavioral rules. The Karpathy skill targets three failure patterns: agents making silent wrong assumptions, over-engineering that turns 50 lines into 500, and orthogonal changes that modify code the agent was never supposed to touch.

Real Skills in This Space

The Divorced Dad Benchmark

From Ponytail's real benchmarks - five everyday tasks, three models:

Use It When

You don't need pretty, you need it live by tonight. Prototypes. MVPs. Hackathons. Internal tools. Anything where working beats perfect and perfect was never shipping anyway.

Install

Create .claude/skills/divorced-dad/SKILL.md:

---
description: "Build the simplest working version first - no over-engineering"
---
# Divorced Dad

Duct tape mode. Ship it.

Before writing ANY code, check this ladder:
1. Does this need to exist at all? → If no, skip it (YAGNI)
2. Does stdlib/native platform handle it? → Use it
3. Is there already an installed dependency? → Use it
4. Can you do it in one line? → Do it in one line
5. Only then: write the minimum code that works

Rules:
- No wrapper classes unless there are 3+ consumers
- No abstraction layers for things used once
- No utility libraries for single functions
- No premature optimization
- No config files for things with sensible defaults
- No error types beyond what the runtime provides
- Ship first. Refactor when you actually need to.

Every shortcut gets a comment: `// divorced-dad: [upgrade path]`
so "later" has a trail when later actually arrives.

Security, accessibility, and data-loss prevention are NEVER on
the chopping block. Cut engineering theater, not safety.

5. Redneck Engineer - /redneck-engineer

What it does: Turns a vague idea into a clear build plan. You get a short PRD, a task list, user flows, constraints, and acceptance criteria - so Claude Code knows exactly what to build instead of wasting hours guessing.

The Problem

You know what you want in your head. You type "build me a dashboard." Claude builds a dashboard - the wrong one. You correct it. It builds a different wrong one. Three hours later you have a Frankenstein dashboard that doesn't match what you imagined.

The issue isn't Claude's coding ability. It's that you never defined what "dashboard" means. Without constraints, acceptance criteria, and user flows, the agent fills every gap with a confident guess.

Real Skills That Solve This

The PRD-generation ecosystem has exploded:

What prd-taskmaster Generates

📄 PRD Created: .taskmaster/docs/prd.md
📊 Overview:
  - Feature: Two-Factor Authentication
  - Complexity: Medium
  - Estimated Effort: 26 tasks, ~119 hours

🎯 Key Requirements:
  1. REQ-001: TOTP/SMS 2FA support
  2. REQ-002: Backup codes for recovery
  3. REQ-003: Login flow integration

⚠️ Quality Validation: 58/60 (EXCELLENT)
  ✅ All required elements present

It asks 12+ detailed questions before generating anything: what problem you're solving, who's the user, what's the tech stack, success metrics, timeline, constraints. Then it scans your codebase for integration points and generates a PRD that both your team and AI tools can execute against.

Use It When

You know what you want in your head but Claude keeps building the wrong thing. Before any feature that touches more than 3 files. Before any build that will take more than 30 minutes.

Install

Create .claude/skills/redneck-engineer/SKILL.md:

---
description: "Turn a vague idea into a clear build plan with PRD, tasks, and acceptance criteria"
---
# Redneck Engineer

Before writing a single line of code, interview the user:

1. What problem are you solving? (Not what feature - what PROBLEM)
2. Who is the user? (Role, context, what they're doing when they need this)
3. What does success look like? (Measurable - numbers, not feelings)
4. What's the tech stack? (Languages, frameworks, databases, infra)
5. What are the constraints? (Time, budget, existing code, compatibility)
6. What is explicitly OUT of scope?

Then produce:
- **PRD** (1 page max): Problem, solution, success metrics, constraints
- **User flows**: Step-by-step what the user does (not what the code does)
- **Task list**: Ordered, with dependencies marked, complexity estimated
- **Acceptance criteria**: Binary pass/fail for each requirement
- **File list**: Which files will be created or modified

Format: Write all output to `docs/PRD.md` in the project root.

Do NOT start coding until the user approves the plan.
Ask "Does this match what you had in mind?" and wait.

The Install Summary

One-Line Setup (All Five)

Create the skill directories and drop each SKILL.md:

mkdir -p ~/.claude/skills/{chinese-grandpa,weaponized-autism,meth-lab,divorced-dad,redneck-engineer}

Then create each SKILL.md in its directory using the templates above.

Trigger

Type the slash command in any Claude Code session:

/chinese-grandpa     # Activate token compression
/weaponized-autism   # Activate exhaustive mode
/meth-lab            # Optimize whatever you hand it
/divorced-dad        # Ship the minimum that works
/redneck-engineer    # Plan before you build

Why "Illegal"?

These names would never pass a corporate review process. But the skills themselves are just SKILL.md files - text documents that change how Claude behaves. There's nothing technically forbidden about any of them. They solve real problems that real developers face every day.

The names are memorable because they're honest. Chinese Grandpa compresses tokens the way an old-school Mandarin speaker compresses language. Weaponized Autism channels the kind of obsessive focus that catches bugs others miss. Meth Lab strips systems down to pure function. Divorced Dad builds with duct tape because the kids need dinner tonight. Redneck Engineer plans with common sense, not corporate frameworks.

Anthropic's marketplace has naming guidelines. These names don't pass them. The skills themselves? They're some of the most useful behavioral modifications you can make to a coding agent.

Follow AI Adventure YT for more AI breakdowns and guides.