Prompt Engineering: Get Stable Output from Models

What you're probably confused about right now

"I wrote a really long prompt. Why are the results still inconsistent?"

Length doesn't equal quality. Structure and verifiability -- that's what matters.

One-line definition

Prompt Engineering is breaking a task into a structured instruction with role/task/context/constraints/output.

Real-life analogy

Give an intern a clear template and rubric, and the results will be way more consistent than "just wing it."

Minimal working example

Role: Python engineer
Task: Clean an email list
Constraints: Deduplicate, preserve order, return invalid items separately
Output: JSON(valid, invalid, count)

Quick quiz (5 min)

1. Rewrite a weak prompt into a structured prompt.
2. Add edge case constraints.
3. Define a fixed output schema.

Quiz answer guide & grading criteria

• Answer direction: write runnable code that covers the core requirements and edge cases from the prompt.
• Criterion 1 (Correctness): Main flow produces correct results, key branches execute.
• Criterion 2 (Readability): Clear variable names, no excessive nesting.
• Criterion 3 (Robustness): Basic protection against null values, type errors, or unexpected input.

Take-home task

Design a prompt for "order number extraction + masking" and iterate through 2 rounds.

Acceptance criteria

You can independently:

• Write structured prompts
• Constrain output format
• Improve stability through iteration

Common errors & debugging steps (beginner edition)

• Can't read the error: start from the last line -- find the error type (TypeError, NameError, etc.), then trace back to the line in your code.
• Not sure about a variable's value: throw in a temporary print(var, type(var)) at key points to verify data looks right.
• Changed code but nothing happened: make sure the file is saved, you're running the right file, and your terminal is in the correct venv.

Common misconceptions

• Misconception: longer prompts are better.
• Reality: clear structure and testability beat length every time.