Evaluation and Iteration
Why One Run Is Not Enough
Understand that model outputs can vary and prompt quality should be judged on representative examples rather than one response.
In this lesson
Why this matters
One successful run does not prove consistency. Test representative normal, long, incomplete, and edge cases, and keep a stable set for comparing versions.
How to apply it
- Define the task and desired outcome.
- Add only context or rules that affect success.
- Specify an output form or checkable criterion when needed.
- Test on realistic input and revise when a failure pattern appears.
Practical example
Weak example:
Weak version of the same task: نجح مثال واحد، إذًا جاهز.The weak version leaves an important part of the task or its boundaries to guesswork. The improved version adds information tied directly to success without treating length or complexity as a goal.
Improved example:
Improved version: اختبر 12 حالة: 6 عادية، 2 طويلة، 2 ناقصة، 2 حدّية؛ سجّل النجاح والفشل حسب المعايير.Short practice
Apply the lesson concept to this task: أنشئ 5 حالات اختبار لمهمة تلخيص متنوعة.
Check your understanding
Check: لماذا لا تكفي تجربة واحدة؟
Answer: لأن الأداء يختلف بين المدخلات ولا تكشف حالة واحدة الاتساق.
Summary
Keep this concept as one tool rather than a magic formula. Tie it to the goal and success criteria, then build on it in the next lesson.