An Agent Skill that roasts AI-hype delusions - then shows the way out.

Works with any agent that supports the open SKILL.md standard: Claude, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, and many more. Can also be used as a plain system prompt with any LLM.

"Clone Facebook in 1 hour." "Rebuild X in 30 minutes." "No need for engineers, AI does everything."

fomo-cc is for the founders and vibe coders who believe this. It roasts them hard in their own language, backs every insult with real numbers, and ends with a plan they can actually ship.

One hour? Are you out of your damn mind? Meta has ~79,000 employees, 20+ years of operations and ~3 billion users on Facebook alone. You have sixty minutes and a prompt. What you launch tonight isn't Facebook - it's a post form running on localhost.

A roast without facts is just noise, and people argue back. A roast with facts leaves nothing to argue with. That's when they actually wake up.

So every roast in fomo-cc does three things:

1. Slaps - hits the specific delusion, with the user's own ridiculous deadline next to the real one
2. Proves - team size, years in operation, infrastructure, the hidden part of the iceberg, and the laws of the target market
3. Redirects - what's realistic in their timeframe, in 1-2 weeks, and in 3-6 months, plus one core loop to build first

• Multilingual, native tone - replies in the user's language and roasts the way a native speaker would. Tone guides for English, Vietnamese, Japanese, Spanish, Chinese and Korean, plus general rules for any other language
• Real numbers only - figures come from annual filings and press, stored in references/. The skill is told never to invent specific numbers
• Market-aware legal points - compliance follows the target market, not the language (e.g. an English prompt about an app for Vietnam gets Vietnamese law)
• Calibrated harshness - full force for "clone Uber this weekend", light touch for "landing page in 2 hours". It won't invent reasons to roast a realistic plan
• Attacks the idea, never the person - no insults about family, appearance, gender, ethnicity, religion or disability

fomo-cc follows the Agent Skills specification: a folder with a SKILL.md and a references/ directory, nothing agent-specific. Copy the fomo-cc/ folder into the skills directory your agent reads.

.agents/skills/ is read by several agents at once (Codex, Gemini CLI, Cursor, GitHub Copilot), so one copy covers them all:

# Personal (all projects)
mkdir -p ~/.agents/skills && cp -r fomo-cc ~/.agents/skills/

# Or project-level (commit it with your repo)
mkdir -p .agents/skills && cp -r fomo-cc .agents/skills/

Other skills-compatible agents (Goose, OpenCode, Amp, Kiro, Roo Code, Junie, ...) have their own paths - see each agent's docs via the Agent Skills client list. Keep the folder name fomo-cc: the spec requires it to match the name field in SKILL.md, otherwise the skill silently fails to load.

Open fomo-cc.skill in the app and click Save skill.

Chat apps without skill support (custom GPTs, Gemini Gems, plain API calls) can still use it. Concatenate the instructions and references into one system prompt:

cat fomo-cc/SKILL.md fomo-cc/references/*.md > fomo-cc-system-prompt.md

Paste the result into the system prompt / custom instructions. Remove the YAML frontmatter at the top if the tool complains. The trade-off: everything loads on every message, instead of only when a roast is requested.

The skill only triggers when you explicitly ask to be roasted. It stays out of the way when you ask for help building something in good faith.

Or call it directly in agents that expose skills as slash commands (Claude Code, GitHub Copilot, ...): /fomo-cc <your idea>.

Since activation depends on each agent matching your request to the skill's description, trigger reliability can vary between agents and models. If it doesn't kick in, name it explicitly ("use the fomo-cc skill").

Every roast follows the same order, with headers in the user's language:

1. Opening slap - 2-4 sentences
2. The hidden iceberg - 5-8 items specific to the product, with real numbers
3. What AI can and can't do - two sides, short
4. The way out - your timeframe → 1-2 weeks → 3-6 months, one core loop
5. Bottom line - one memorable closing line

Target length: about 300-600 words. No emoji.

fomo-cc/
├── SKILL.md                         # Instructions, tone rules, output structure
├── references/
│   ├── reality-numbers.md           # Headcount, years, users for Meta, X, Uber, Grab, Shopee, Netflix...
│   ├── iceberg.md                   # Hidden engineering, security, legal and business work
│   └── tone-by-language.md          # How to roast natively in each language
└── evals/
    └── evals.json                   # Test prompts

Numbers in references/reality-numbers.md were checked in September 2026. Headcounts and user counts change every quarter. When the skill runs with web search, it can look up products that aren't in the file; otherwise it falls back to orders of magnitude.

• Add a product - append to references/reality-numbers.md with a source (annual report, 10-K/20-F, reputable press). No source, no number
• Add a language - add a section to references/tone-by-language.md: register/pronouns, everyday swear words (mild, no slurs), idioms for "get real", translated headers. If you're not a native speaker, get one to review it
• Add a market's laws - extend the legal section in references/iceberg.md, grouped by market
• Test your change - add a prompt to evals/evals.json and compare the output before and after

This skill swears on purpose. Use it on yourself, or on people who asked for it. Legal and compliance points are pointers for a reality check, not legal advice.