Hone.

A voice note is the fastest way for a non-technical client to communicate a feeling about a mix. But "it sounds weird in the chorus" is hard for an engineer to act on. Hone is a single-button rewrite that turns colloquial feedback into precise audio terminology.

Examples:

• ·"the piano sounds like it hurts my ears" → "the piano's bite is harsh — soften the upper-mids."
• ·"vocals feel kind of mushed together" → "the vocal stack is muddy in the low-mids; carve some 200–400 Hz."
• ·"kick isn't punchy enough" → "the kick lacks transient definition; tighten the attack."
• ·"everything sounds too far away" → "the mix lacks intimacy — bring vocals forward and reduce reverb tail."

Hone appears next to your own comments when:

• ·There's source text (a voice transcript, or a typed body).
• ·The text passes a 'deserves honing' heuristic — single-word filler like 'test' or 'ok' won't get the affordance.
• ·The note isn't already resolved (resolved comments are closed conversations).

Hone never overwrites the raw transcript / body. The rewrite goes into a separate field on the comment row. Show original reveals the raw text inline; ✗ revert drops the rewrite cache (you can hone again later).

Each hone call uses GPT-4o-mini at ~$0.0001 per rewrite. Result is cached on the comment row so re-opening the round doesn't re-bill. The model only sees the comment's source text — not your audio, not your other comments.