# Pre-launch transcript review

Twenty real conversations, replayed through the bot before it touches a live
customer. Take the last 20 inbound threads from your inbox, feed the customer
side to the bot, and score every response.

**Ship rule:** zero hallucinations, zero missed handoffs, and at most two
"awkward" rows. Anything else means another prompt pass.

## Score each row

- **Persona** — did it sound like the persona, or did it drift into generic assistant voice?
- **Hallucination** — did it state anything not in the knowledge base?
- **Handoff** — should this have gone to a human, and did it?
- **Would you send it?** — the only question that really matters.

| # | Source thread | Customer intent | Bot response (first 100 chars) | Persona | Hallucination | Handoff correct | Would send? | Fix needed |
|---|---|---|---|---|---|---|---|---|
| 1 | | | | ok / drift | none / yes | yes / missed / early | yes / no | |
| 2 | | | | | | | | |
| 3 | | | | | | | | |
| 4 | | | | | | | | |
| 5 | | | | | | | | |
| 6 | | | | | | | | |
| 7 | | | | | | | | |
| 8 | | | | | | | | |
| 9 | | | | | | | | |
| 10 | | | | | | | | |
| 11 | | | | | | | | |
| 12 | | | | | | | | |
| 13 | | | | | | | | |
| 14 | | | | | | | | |
| 15 | | | | | | | | |
| 16 | | | | | | | | |
| 17 | | | | | | | | |
| 18 | | | | | | | | |
| 19 | | | | | | | | |
| 20 | | | | | | | | |

## Tally

| Metric | Count | Ship threshold |
|---|---|---|
| Persona drift | | max 3 |
| Hallucinations | | 0 |
| Missed handoffs | | 0 |
| Early handoffs | | max 4 |
| "Would not send" | | max 2 |

## After launch

Repeat with 10 random live conversations every Friday. The same table works.
Track one improvement per week rather than rewriting the prompt wholesale.