A practical checklist distilled from this entire section — the habits that separate reliable, production-grade prompts from ones that happen to work in a quick demo.
Clarity & Structure
- Separate instructions, context/data, and the actual task into clearly identifiable sections (see Prompt Structure)
- Be explicit about output format, length, and tone rather than leaving it to the model's default judgment
- Use delimiters to clearly mark untrusted/external content
Technique Selection
- Start with zero-shot; add examples (few-shot) only if output consistency actually requires it
- Reserve chain-of-thought/reasoning techniques for genuinely multi-step tasks, not simple lookups
- Decompose complex tasks into a chain of smaller, testable prompts when a single prompt's reliability degrades
Engineering Discipline
- Use templates for any prompt pattern reused across requests — don't hand-write similar prompts ad hoc throughout a codebase
- Maintain a real evaluation set with edge cases, and re-run it before shipping any prompt change (see Prompt Evaluation)
- Change one thing at a time when optimizing, so you know what actually caused an improvement or regression
- Version-control prompts like any other piece of application logic
Security & Safety
- Treat all external/retrieved content as untrusted input (see Prompt Injection)
- Never let raw model output execute as code/SQL/commands without validation
- Don't rely on the system prompt alone as a security boundary — back critical constraints with real validation and least-privilege access
Cost & Performance
- Keep prompts as token-efficient as the task allows — cut filler, don't over-include context "just in case" (see Token Optimization)
- Match model size to task complexity rather than defaulting to the most capable (and expensive) model for everything
Practical Use Case
Use this list as a review checklist before shipping any new LLM feature to production — most real-world prompt failures trace back to skipping one of these basics (no evaluation set, no format specification, untrusted content treated as trusted).
Common Mistakes
- Treating this list as a one-time setup step rather than an ongoing practice — prompts need revisiting as usage patterns, models, and requirements evolve
- Applying every technique from this section to every prompt regardless of whether the task actually calls for it — added complexity without benefit
Interview Relevance
"Walk me through your checklist for reviewing a prompt before it goes to production" — structure, technique fit, evaluation, security, and cost are the five categories a comprehensive answer should touch.
Practice Question
Review this prompt against the checklist above and list at least three concrete improvements: "Answer the user's question using the info: [context] Question: [question]"