The two-draft split is the right instinct. The failure mode you're describing — a model confidently inventing a price or scope line nobody agreed to — is exactly why "summarize the call" isn't the same job as "write the proposal." Curious whether you've run into models still leaking the internal draft's assumptions into the client-facing one when the boundary between the two isn't hard-coded.
Fifteen minutes of judgment beats a full day of writing every time.
The best system starts with clear inputs.
The two-draft split is the right instinct. The failure mode you're describing — a model confidently inventing a price or scope line nobody agreed to — is exactly why "summarize the call" isn't the same job as "write the proposal." Curious whether you've run into models still leaking the internal draft's assumptions into the client-facing one when the boundary between the two isn't hard-coded.
The real value is getting repetitive work out of the way, so human judgment stays focused where it matters.