When I hand a card news draft to one of the Dream Team bots, I list a few check items inside the prompt, such as 「title within 20 characters」, 「exactly 5 hashtags」, and 「verify Traditional Chinese characters」. Once the list reached about ten items, one or two always went missing. So I ran a small experiment where the only thing I changed was where the checklist sits.
How the test worked
Same topic, same model, same 8-item checklist, two placements.
• A: checklist at the very top of the prompt
• B: checklist at the very bottom of the prompt
I collected 10 drafts for each and logged in a table how many of the 8 items were respected.
Results
A (top): in 6 of 10 runs, at least one item was missed
B (bottom): in 2 of 10 runs, at least one item was missed
The sample is tiny, so this is not a scientific conclusion. Still, in our setup (a Windows device + a Mac mini + Claude Code), the direction was fairly clear. At the moment the bot finishes reading and starts writing, it seems to follow the sentences closest to it most faithfully.
How to try it yourself
1. Open your prompt file.
2. Gather the check items into one block and move it to the very end.
3. Add one line above it: 「Before you answer, check each item below one by one.」
4. Number the items and keep them short.
(main instructions)
...
Final check
1. Title within 20 characters
2. 5 hashtags
3. Traditional characters only
One more note
Too many items weakened the effect. Around 8 worked well, and at 12 the misses started coming back. The whole test takes about 10 minutes, so if one of your bots keeps dropping items, start by changing the position.