novel-characters

Point the skill at your novel text file and it runs a two-pass pipeline: scanning overlapping chunks for characters and quotes, then building full profile cards for the top characters by prominence. Output includes cast.json, a Markdown report, and a self-contained report.html.

The merge step indexes characters by both name and alias, so different forms of address across chunks converge onto one person. Containment-based merge candidates that exact matching cannot resolve are listed for model review and applied deterministically from a merges.json file.

Yes, pass --lang en or --lang ja; Chinese, English, and Japanese UI strings are built in. Other languages also work because the skill translates UI labels on the fly and stores them in cast.json, while image and TTS prompts always stay in English.

Yes, codex is only needed for the optional model-sheet image generation step. Without it, the skill still delivers the full cast.json, prompts, and reports; only the per-character sheet images are skipped.

A single run caps at 24 chunks of 40,000 characters with overlap, covering roughly 930,000 characters net. Beyond that limit the skill reports truncated explicitly rather than silently dropping the tail of the text.

The validator enforces hard rules: evidence quotes must be verbatim contiguous spans of the source, image prompts must not contain character names, human-readable fields must follow the report language while machine prompts stay English, and the style preset must match its negative prompt.