Review summary with quotes: pros and cons backed by review IDs

What it does

Summarizes customer reviews with evidence instead of guesswork. It never invents reviews or quotes.

• Reads CSV, TSV, JSON, JSONL or plain text (blank-line separated) and detects ID, rating, text and date columns (assigns R1, R2... when IDs are missing)
• Rating distribution and mean, plus positive / mixed / negative counts
• Counts 14 themes (price, quality, defects, size, ease of use, noise, delivery, instructions, support and more) with Japanese and English keyword lists, listing review IDs and verbatim sentence quotes per theme
• Verifies that every "quote" [ID] in the finished summary really exists in the cited review and fails loudly on missing quotes or unknown IDs

How to use

python scripts/review_themes.py analyze reviews.csv --out themes.json --md themes.md
python scripts/review_themes.py verify summary.md --reviews reviews.csv

Copy references/themes.json, add product-specific keywords and pass it with --themes. Samples are in examples/.

Requirements

• Python 3 (standard library only)
• An agent such as Claude, Codex, Hermes Agent or Cursor

Caveats

• Keyword counting misses paraphrases and can count negations ("not noisy"); the agent reads the quotes before concluding.
• It does not scrape reviews; use data you are allowed to use.
• Reviewer personal data is kept out of the summary. Follow each platform's rules when quoting publicly.