immortal-skill

Export chat records from platforms like WeChat, Feishu, or iMessage, then run the collector CLI to normalize them into a Markdown corpus. The Skill extracts interaction style, memory, personality, and procedural knowledge into separate files with evidence grading, producing a loadable Skill folder.

Supported sources include Feishu, DingTalk, Slack, Discord, Telegram, and Email via APIs, WeChat and iMessage via local databases, and WhatsApp, Twitter/X, and Google Takeout via archive files. Any plain text, JSON, CSV, or Markdown file can also be imported manually.

Yes, the public-figure persona template supports this using only public, traceable sources such as talks, interviews, blogs, and podcasts. The output represents their publicly demonstrated reasoning frameworks as a reference advisor, not an impersonation of the real person.

The family persona template requires confirming whether other living family members should be informed before distilling a deceased relative. Outputs must be labeled as memory aids based on limited material, not complete reproductions of the person.

Contradictions are never force-merged. Conflicting claims from different sources or time periods are written to a dedicated conflicts.md file, preserving both versions with their evidence levels so the persona reflects real human inconsistency.

This approach uses structured extraction rather than fine-tuning because personal chat data is usually too sparse for quality training. Extraction offers better interpretability and editability, but the resulting persona depends heavily on the volume and coverage of source material.