AI Setup for Humor: Config and 7 Languages

How to Configure AI for Humor Generation: Experience Building a Multilingual Data Site

A developer created the absurd statistical site scratchstats.fun — a real-time global counter of 'male scratching' with a map and UN WPP 2024 data. The project, with a Go backend and D3.js frontend, was entirely generated by Claude (Opus 4.6). The key insight: humor in AI is achieved not by randomness, but through strict prompt configuration with tone, contrast, and cultural adaptation across 7 languages.

The lead gets straight to the point: instead of cliché jokes, Claude delivered content with academic seriousness about a trivial topic. The cookie banner, FAQ, disclaimer — each element maintains contrast: data presented like a Bloomberg report, the topic treated like WHO statistics. The project took 11 hours of pure work, stretched due to API limits.

Technical Stack and Specification

Backend in Go: event-driven architecture, high availability. A sinusoidal wakefulness model across 10 regions, counters, world map, tables. Frontend (D3.js, animations, layout) — 100% Claude based on specs: 8 user stories, 36 FR, edge cases.

Work in Cowork + Claude Code. Cowork for ideas, tone, strategy (project manager, co-writer roles). Claude Code — only code and translations. A CLAUDE.md file with a Humor & Tone Guide — iteratively refined through dialogue.

Example cookie banner (generated without a humor request):

Advertising cookies:

Imagine: targeted ads for premium underwear, anti-itch cream, and motivational apps. We've spared you that. Don't thank us.

Social cookies:

We could have added a 'Share on LinkedIn' button so you could post your visit to a testicle-scratching statistics site on a professional network. We decided against it. Your career is safe.

Google AdInline article slot

Disclaimer:

This site does not encourage scratching. We recommend channeling that energy into something productive. Free your hands for great deeds.

Humor Guide: Engineering Humor

Humor as config: the contrast between seriousness and absurdity. Universal rules:

• Unflappable academic delivery — data treated like GDP.
• Never apologize for the topic — legitimate research.
• Self-aware institutional voice.
• Parenthetical remarks for cultural commentary.
• No emojis, terminal aesthetic.
• Laugh with, not at.

Voice by language (7 JSON files, ~450 keys each):

• EN: Academic absurdism, BBC tone.
• RU: Direct, cheeky, no euphemisms.
• UK: Direct + warm, its own voice (not a copy of RU).
• DE: Bureaucracy, 'Ordnung muss sein', compound words.
• ES: Conversational, siesta motifs.
• CS: Švejkian absurdity, beer.
• BG: Folklore, rakia, proverbs.

Anti-patterns:

• Don't break the fourth wall.
• No toilet humor.
• Cultural adaptation > literal translation.
• Unique comedic register per language.
• Directness for RU/CS/BG.
• EN — formal.

Example FAQ "What if everyone stopped scratching?":

Global productivity would increase by 45.8 billion seconds/day (2.645 billion men × 3 sessions × 5 sec). That's 1453 years of labor daily. Enough for 2.3 ISS/year.

Numbers are verifiable, the conclusion is trolling.

Multilingual Adaptation

Not translation, but localization per guide. Tone discussion before generation. Verification: RU/UK/BG/EN — author, others — colleagues from those countries.

Comparison of FAQ "Scratching right now?":

• RU: "No. But statistically you are one of 400,000. You are not special. A data point."
• UK: "We know. You are one of 400,000. You are the reason for this site."

Different registers: RU — cold, UK — empathetic. Bulgarian press kit with a grandma and banitsa — pure folklore.

Belarusian was rejected: AI honestly pointed out weakness in cultural nuances.

Press Kit and Team Dynamics

Claude independently proposed a press kit: boilerplate, quotes, headlines.

Quotes:

"The hardest part wasn't the math. Explaining it to family." — Head of Research. "Pivot to nose-picking: TAM is larger (91% of adults), stigma is worse." — Strategy Director.

TAM analysis: nose universal, but taboo. Joke at pitch level.

Cowork + Claude: role separation, iterations. Author (53 years old, Go expert) — oversight.

Key Takeaways

• Humor in AI — prompt configuration: tone, contrast, anti-patterns.
• Cultural adaptation > translation: 7 languages with unique voices.
• Model honesty: rejecting weak languages.
• Team workflow: Cowork (ideas), Claude Code (implementation).
• Absurd data site as a testbed for humor engineering.

— Editorial Team