Prompt Engineering
Topic
Name: Prompt engineering
Why it matters for FDE roles: Customer-facing AI work often starts with turning an ambiguous human task into reliable model instructions, then improving those instructions through examples, tests, and workflow constraints.
Plain-English Definition
Prompt engineering is the practice of giving an AI system the right task, context, constraints, examples, and output format so it can produce useful results repeatedly.
Where It Shows Up
- Job listing signal: LLM workflows, prompt design, AI product implementation, copilots, evaluation, structured outputs.
- Portfolio project connection: Ops Knowledge Copilot needs prompts for summaries, next-action extraction, uncertainty flags, and cited recommendations.
- Real customer scenario: A customer wants an AI assistant to turn messy onboarding notes into a clean status update without inventing missing facts.
Core Concepts
- Task framing: define the job the model is doing.
- Context selection: include only the information needed for the task.
- Output constraints: require JSON, citations, labels, or short sections when software needs predictable results.
- Few-shot examples: show good and bad outputs for recurring tasks.
- Iteration with tests: improve prompts against a small set of representative examples.
- System boundaries: tell the model what it may not do and when to ask for review.
Failure Modes
- Stuffing too much context into a prompt instead of selecting relevant inputs.
- Asking for vague judgment without defining criteria.
- Depending on prompt text alone for safety instead of using validation and permissions.
- No regression examples, so a prompt change silently breaks old cases.
- Outputs that sound confident but do not expose sources or uncertainty.
Tiny Practice Task
Write a prompt that turns one operational record into {summary, next_action, owner_guess, urgency, missing_information, citations} and test it on three intentionally messy examples.
Interview Language
One sentence I could say in an interview:
I treat prompts as part of the product surface: they need clear task framing, controlled context, structured outputs, examples, and regression tests just like other implementation logic.