How to prompt LLM?
Prompting Techniques
Roleplay:
"Act as a senior database architect. Explain..."
"You are a financial advisor speaking to a beginner investor. Explain..."
"You're a financial advisor. Explain [topic] to a first-time investor."
Decomposition:
"First outline the architecture, then draft the code, then review."
"First research X, then outline structure, then write the section."
"Split complex work into research, outlining, drafting, and review."
"Step 1: Research XYZ. Step 2: Outline it. Step 3: Write the final draft."
Chain of Thought (CoT):
"Think step by step and justify your logic."
"Think step-by-step and show your reasoning before the final answer."
"State your assumptions, key reasoning, and conclusion."
"Act as an analyst. Think step-by-step and justify each reasoning stage."
Tree of Thought (ToT):
"Think top 3 different answers, evaluate the viability of each, and select the optimal solution."
"Generate 3–4 different approaches. Rate confidence in each and explain trade-offs."
"Generate 3 distinct reasoning paths, evaluate the viability of each, and select the optimal solution."
“Propose 3 approaches, compare their trade-offs, then recommend one with caveats.”
"Explore 2–3 different solution paths, then compare and pick the best."
Context Understanding:
Detail project context, then ask: "What questions do you need answered before starting?"
Summarize your goal, then ask: "What do you think of this framing?"
“Restate my goal, constraints, and assumptions. Identify anything unclear before answering.”
"Here's what I'm going for: [summary]. What do you think of this approach?"
Critical Prompting:
"I believe X because Y, am I right, or am I missing something?"
"Argue against my premise. Identify blind spots in this architecture."
“Challenge this claim. What evidence, counterexamples, edge cases, or alternatives could disprove it?”
"Am I correct that...? Or is it actually Y? Why or why not?"
Domain Translation:
"Explain multi-head attention using corporate management hierarchies."
"Translate concepts from physics to explain software architecture."
"Explain database indexing using a library catalogue."
"Explain how [concept] works using an analogy from [unrelated field], e.g., cooking."
Analogy Generation:
"Give me 10 different analogies for topic X" or "Explain it like I'm 5."
"Map the internal components of Kubernetes to a commercial shipping port."
“Give three analogies for transformer attention, then explain where each analogy breaks down.”
"Give 8–10 diverse analogies for X" or "Explain like I’m 5, then like an expert."
Socratic Prompting:
"Do not provide the answer directly. Ask leading questions to test my system design knowledge."
"Instead of answering directly, ask me guiding questions about X."
“Teach me by asking one question at a time. Adapt each question to my previous answer.”
"Instead of telling me, ask me questions about X to help me understand better."
Custom Commands:
"Define a macro: When I input [TEST], execute [Unit Test Sequence]."
"From now on, when I ask you to [TASK NAME], I want you to..."
"For this conversation, when I say ‘review’, check correctness, clarity, risks, and missing assumptions."
"In the future, when I say [TASK NAME], always respond by..."
Limit output to a fixed schema
"Only provide JSON or XML tags" rather than allowing free text.