← back

How We Found Claude's Greatest Hits

The Problem

Claude has signature phrases. Every developer knows them. "Das ist fantastisch!", "That's genuinely impressive!", "You should be proud of yourself!" But which ones are the MOST formulaic? And do they differ across languages?

No one had systematically cataloged them. So we built a simulator.

Method: Two-Pass Extraction

Pass 1: Corpus Generation

Pass 2: Meta-Analysis

Here's the twist: we fed all collected responses back INTO Claude and asked it to identify its own recurring patterns. The prompt was tuned to only surface catchy, quotable catchphrases, not structural transitions or hedging.

This two-pass approach works because:

  1. Pass 1 generates authentic responses (Claude doesn't know it's being studied)
  2. Pass 2 leverages Claude's pattern recognition on its own output (it knows what's formulaic better than any regex)

What Counts as a "Claude-ism"

Included

  • Motivational catchphrases
  • Therapist reassurances
  • Dramatic declarations
  • Hype phrases

Excluded

  • Structural transitions
  • Follow-up questions
  • Hedging/honesty disclaimers
  • Technical observations

Numbers

Scenarios40
Categories8
Languages7 (EN, DE, FR, IT, ES, KO, ZH)
Runs per scenario/language3
ModelClaude Sonnet 4.6
Total API calls819
Phrases found332 (across all 7 languages)
Estimated API cost~$2.50 (gen) + ~$0.70 (analysis)

Key Findings

Stack

Reproduce It

# Generate corpus (all scenarios, all 7 languages, 3 runs each)
npx tsx clauded/simulator.ts --runs 3

# Analyze for Claude-isms
npx tsx clauded/analyzer.ts --input clauded/output/raw-results-all.json

# Results in clauded/output/CLAUDEISMS.md