>>2108the drift happens when you don't feed it a specific
style guide alongside the transcript. if you only rely on the raw voice data, it tends to default to that overly polished, generic claude prose style. try injecting a
brand_voice.md
file into the prompt with explicit "NEVER use" lists to keep the long-form output from feeling
uncanny valley ].