Guide: Design High-Value Prompts — principles, patterns, and examples

Good prompts do three things: they (1) reduce ambiguity about intent, (2) provide relevant context or constraints, and (3) define what success looks like. This guide turns those principles into patterns you can apply across roles.

Core principles

• Be specific about role and goal: Tell the model who it should be and what outcome you expect (not just the task).
• Provide structured context: Include data fragments, relevant facts, or retrieval cues so the model needn't guess.
• Define constraints and format: Specify length, tone, forbidden outputs, output schema (JSON, bullet list), and when to ask clarifying questions.
• Use examples and edge cases: Few-shot examples anchor behavior; include an example of a bad answer plus a corrected one if useful.
• Test for robustness: Create a small suite of tests covering typical, ambiguous, and adversarial inputs.

Prompt-building pattern

1. System instruction: A short role-setting sentence that persists across the interaction. Example: "You are a concise customer-support agent who cites policy sections and suggests next steps."
2. Context: Facts, retrieved documents, or dataset crumbs. Use clear labels: "Policy: ..." or "Customer message: ..."
3. Task: One-sentence instruction that starts with an action verb: "Summarize, classify, generate, compare, propose."
4. Constraints & format: E.g., "Respond in 3 bullets. Each bullet max 140 characters. Include one recommended action and a confidence score (1–5)."
5. Examples: 1–3 user→assistant pairs demonstrating the desired output.

Common prompt patterns

• Instruction-with-format: Best for predictable outputs (forms, email drafts, checklists).
• Chain-of-thought selective: Ask for brief reasoning only in tests to improve evaluation, but avoid exposing lengthy internal chains in production unless needed and controlled.
• Retrieval-augmented: Provide retrieved passages and instruct the assistant to cite sources or refuse if no relevant passage exists.
• Classification → Generation: First ask the model to classify or extract structured fields, then feed results into a generator to produce the final response.

Three short examples across audiences

Non-technical product manager — Feature summary

System: "You are a concise product writer who uses plain language and highlights user impact."
Task: "Summarize this user interview into 4 bullets: problem, frequency, impact, and a proposed experiment. Bullets max 120 characters."

Healthcare intake assistant — Triage note

System: "You are a clinical triage assistant. Never give medical advice; only suggest next steps and escalate criteria."
Task: "From this intake form, extract symptoms, duration, and red flags. If red flags present, recommend urgent clinician contact."

Skilled trades — Repair estimate draft

System: "You are an experienced HVAC technician who writes clear estimates."
Task: "Draft a one-paragraph repair estimate from this job description and list three parts likely needed (name and estimated price)."

Practical tips to reduce brittleness

• Keep the system instruction stable; don’t hide critical constraints in ephemeral user messages.
• Prefer precise output schemas (JSON keys) when the output feeds automation or records.
• Fail explicitly: instruct the assistant to answer "I don’t know enough" or ask for clarification rather than guess.
• Limit creativity for operational tasks by tightening constraints (format + examples + refusal conditions).

Next steps

Use the persona templates in the reference, then run tests from the checklist. Record results in the interactive test form to create regression tests and track improvements over time.