Manus to check ryokan details across language pages

Pon tested Manus and Claude Code on a detailed comparison of three Kyushu hot spring inns. The task covered twelve criteria per property, including rooms, private baths, meals, transfers, children and pets. Each cell needed a source, and unknown information could remain unknown.

Claude Code orchestrated the Manus request through its API while conducting the same research itself. Manus searched English official pages and booking screens that Claude's Japanese page search missed. That surfaced transfer information, check in times and child pricing categories.

More filled cells did not mean more confirmed facts. The account reported thirty one populated cells for Manus, with twenty four supported by official pages, compared with twenty eight and twenty six for Claude. Manus missed a published room count and supplied directions from a station no longer served by trains. Claude also needed correction when a search summary described dinner in the guest room but the official page said a private dining room.

Pon examined disagreements and later had the agents assess each other's output. They noted biases in the scoring and differences between subscription cost allocation and Manus credits. The result was an actual comparison experiment, not a hotel stay or booking. Its useful output was a sourced matrix whose gaps and contradictions could be inspected.

Prompt
Original prompt

Source: note.com · ポン (@ponberi)