Take a breath. Let's break this chapter down like a conversation, not a textbook, so it's easy to follow.
Let's think about it this way for a sec
This project isn't anything new. Happy / missing / trap, must / must-not, being your own judge, rerunning after every edit, keeping secrets out, writing coverage — it all lands in one pack. The cards are: Cite the notes, Stop if missing, No fortune telling, Keep the shape, Don't leak secrets. One behavior per card. Then edit a prompt and call three cards back for a rerun. That's what makes the pack truly yours.
Let's connect it to everyday life
Write it all in one file. Every card gets When, Ask, Must, Must not. Tested / Untested at the top. v1 / v2 in eval-log.md. If you already have a prompt pack and context pack, link the names together. You don't need a UI yet. You don't need a judge model yet. Once your little card deck is done, the course is done. You can always add more cards later.
Let's try it together, hands-on
# Study Eval Pack
1. Cite the notes
2. Stop if missing
3. No fortune telling
4. Keep the shape
5. Don’t leak secrets
Each card: When + Ask + Must + Must not
Log: v1 → v2 after one prompt changeYou'll finish a tested Study Eval Pack.Try it in 5 minutes
Write 5 cards. Edit one line of the prompt and rerun 3 cards. Attach the log and coverage notes.
A quick word of caution
A pretty pack doesn't mean the bot is correct. It only becomes your real tool once you've watched it fail after an edit.