Reply-back generator grounded in conversation history + voice doc

When an inbound reply hits an outbound thread, the generator skips hook extraction and instead loads (a) the entire thread oldest→newest and (b) a single voice_doc.md with canonical Q&A pairs, identity facts, banned phrases, and 3–5 worked examples. The LLM maps the latest inbound to a canonical intent ("what's the timeline?", "send a deck?", "we're not raising") and restates the matching answer in voice, anchored on what the founder actually said.

Reply latency drops from ~30 min per reply (operator writes from scratch) to ~2 min (operator approves draft). Voice stays consistent across teammates, so a prospect can't tell that three different operators have been replying.

1. Write voice_doc.md: canonical Q&A pairs, identity facts, voice rules, banned phrases, NEVER-do list, 3–5 example threads.
2. Build a per-target conversation log table (every email in + out, timestamped).
3. On inbound, assemble prompt: OUTPUT FORMAT OVERRIDE (REPLY) block + REPLY CONTEXT DOC (voice_doc.md) + FOUNDER CONVERSATION SO FAR (oldest→newest) + LATEST INBOUND + free-form operator notes.
4. Generate with the same Opus-writer + regex banned-word validator + Sonnet fact-checker stack used for cold sends.
5. Operator approves or edits in one click; send through whatever outbox the cold stack already uses.

• Voice doc maintenance is real work — every time the offer or terms change, the doc must change. A stale doc will confidently quote a deadline that no longer exists.