Soft caps (0 = never, 4 = strongly when invited). The three highlighted dials — tension, warmth, directness — render as the explicit balance line in every prompt. Thread context still wins: if she is upset, the model will lower tension / never-escalate even if the dial says otherwise.
Free-form operator directives rendered verbatim into every reply prompt. Up to ~1200 characters. Use this for guardrails that the dials can't capture.
Short descriptors of how you sound. Rendered as "Voice traits:" in the prompt.
Patterns you want the coach to reinforce. Shown to the model as "patterns that land" — biasing, not scripting.
Shapes the coach must refuse (meta/assistant voice, cold pivots on bids, therapy-speak, thin placeholder imperatives, etc.). Keep these sharp — they are the only thing stopping regression to the known failure modes.
Refreshed by the "Re-learn from history" button. Voice samples are real outgoing lines of yours that scored as characteristic + sendable; avoid samples come from outcome memory (lines that landed poorly).