Corpus Persona-Extraction

Distill how a specific voice thinks and speaks from a pile of session logs, and persist it as a machine-usable lexicon that downstream agents can write against.

Why this exists

Agents that work with a long-running human collaborator keep re-learning the same voice from scratch: the coinages, the load-bearing metaphors, the sentence shapes that mean "yes, proceed" versus "you've missed the point." A lexicon file makes that knowledge durable and composable — voice-enforcement gates can score against it, drafting skills can imitate it, and translation layers (persona storefronts) can project from it.

Core doctrine

• User coinages outrank standard vocabulary. When the corpus shows an idiosyncratic term for a concept ("vacuum", "organ", "cartridge", "liturgy"), the lexicon records the coinage as primary and the standard term as gloss — never the reverse.
• Append-only. Each extraction run adds a dated block. Prior blocks are evidence of drift, not errors to clean up.
• Separate speakers ruthlessly. A transcript interleaves human, assistant, and tool text. Only the persona's own turns feed frequency analysis; assistant paraphrase contaminates the signal.

Workflow