The Constraint Stack
Add constraints one at a time until output quality collapses, and find where the ceiling actually is. Teaches more about prompt design in twenty minutes than reading about it for a week.
Why this one
People believe more constraints always mean better output. There is a real limit, it varies by model and task, and the only way to know where it is for yours is to walk into it deliberately.
Steps
Start with one bare instruction
Something like "write a product description for a reusable water bottle". Keep the output. This is your baseline and you will want to compare against it later.
Add constraints one at a time
Audience. Then length. Then tone. Then a banned-words list. Then a required structure. Then a specific opening. Then something to avoid. Re-run after each addition, keeping everything else identical.
Find where it breaks
At some point the model starts dropping constraints, or satisfying them so literally that the writing goes stiff. Note the number of constraints where that begins — for most current models it is somewhere between five and eight.
Try grouping instead of listing
Take the constraints that broke it and express them as two or three grouped rules rather than seven separate ones. Grouped constraints usually survive further, which tells you something about how these models weight instructions.
You should end up with
A documented breaking point and one grouping strategy that pushed past it.
Done when
- You identified the constraint count where output degraded
- You noticed which constraint got dropped first
- Grouping changed the result measurably
If you want to go further
- Run the same ladder on a different model — the ceilings differ noticeably