The 10-point rubric

Every prompt you ship in the capstone gets graded against the same ten criteria. One point per criterion, no half points. Anything below 8/10 means the prompt is not ready — redesign it before claiming it shipped.

The rubric

#	Criterion	Points
1	Role / persona stated	1
2	Specific task with measurable outcome	1
3	Input format defined	1
4	Output format defined (length, structure, tone)	1
5	At least 2 hard constraints (must / must-not)	1
6	At least 1 worked example (input → expected output)	1
7	Refusal or fallback when out-of-scope	1
8	"I don't know" trigger when info is missing	1
9	Tested on at least one real input — output meets spec	1
10	Total length under 400 words (system prompts that don't fit in working memory rot)	1

Passing = 8/10 (80%). Anything below means redesign before shipping. (The 80% pass bar is consistent across all capstones in the Prompt Engineering Path — Code uses 40/50, Cross-Model uses a five-row rubric — same bar, different scales.)

How to grade each item, honestly

The trap with self-grading is being generous with yourself. The criteria are written so that "almost" does not score. Here is how to grade strictly:

• Role stated — there is a single sentence in the prompt that names the role and the brand. "You are helpful" does not count.
• Specific task with measurable outcome — could a teammate write a one-line check that decides if the output passes? If not, the task is not measurable.
• Input format defined — the prompt names what the user message will look like (a paragraph, a JSON, a code block, a transcript). If you would not know how to feed it, fail it.
• Output format defined — length AND structure AND tone are all specified. Two of three fails.
• 2 hard constraints — count them. They must be specific ("never use 'unfortunately'") not vague ("be friendly").
• 1 worked example — a real input and a real expected output, not a description of "what good output looks like".
• Refusal scope — there is a line covering what the assistant will NOT do.
• "I don't know" trigger — there is a line covering what the assistant says when it lacks the info.
• Tested on real input — you actually fed it a real message and read the output, not "I think it would work".
• Under 400 words — open the prompt in a word counter. Be honest.

Why this rubric exists

The rubric is reverse-engineered from the failure modes of prompts that crashed in production. Each criterion maps to a class of bug:

• Missing role → tone drift across replies.
• Vague task → outputs that "look reasonable" but do not solve the user's problem.
• No format → markdown headings everywhere, or none at all.
• No examples → tone matches the rules but the shape is wrong.
• No refusal → the assistant cheerfully answers off-brand questions.
• No "I don't know" trigger → confident hallucination.
• Untested → all of the above shipped at once.
• Over 400 words → contradictions inside the prompt that you cannot see at a glance.

If you score 8 or above on a prompt, you have avoided all eight failure classes. That is the bar.

Next: three example scenarios you can study or use as templates. :::