Code Immortality: Digital Persona Distillation

Problem

Many agents only imitate a person’s tone but cannot reproduce how that person would reason about a hard problem. Code Immortality shifts persona distillation from quote cloning to thinking-structure extraction: it first identifies the target’s mental models, decision heuristics, expression DNA, anti-patterns, and honest limits, then produces a runnable SKILL.md so later conversations can operate within the same cognitive framework.

How It Works

• Entry routing: a clear name enters the direct path; vague needs first get diagnosed and mapped to two or three persona or topic candidates.
• Multi-source collection: six parallel subagents cover writings, conversations, expression, external views, decisions, and timeline; local PDFs, interview transcripts, and subtitles are treated as high-priority first-hand material.
• Model synthesis: recurring claims that are also supported by decisions are abstracted into mental models, while contradictions and controversies are preserved instead of smoothing generic advice into a “unique insight.”
• Validation and refinement: the skill is checked against first-hand source ratio, expression recognizability, and explicit limits, then reviewed by two agents focused on workflow clarity and role-play rules to produce concrete edits.
• Self-evolution: each completed run writes a runtime log; after repeated runs it tunes agent prompts and source policy, and updates existing skills by incrementally refreshing conversations, decisions, and timeline.

Boundaries

It is useful as a thinking advisor, decision reference, or role-assistance layer, not as a substitute for the person in legal, medical, or investment advice. Quality drops when first-hand material is scarce, topics are fast-moving, or the persona is obscure; Zhihu, WeChat public accounts, and encyclopedic sources are excluded, so low-profile figures often benefit from user-provided local corpus.