Every advanced prompt is built from these primitives. Learn them once, apply them forever.
Assign the model a specific expert identity to shift its knowledge retrieval and tone.
You are a {role} with {years} years of experience in {domain}. Your task is to {task}. Constraints: {constraints}.Priming with a role activates domain-specific vocabulary and reasoning patterns latent in training data.
Force the model to reason step-by-step before answering — reduces logic errors dramatically.
Think through this step by step inside <thinking></thinking> tags. Only after finishing your reasoning, give the final answer inside <answer></answer>.
Explicit intermediate reasoning is the single largest accuracy booster for math, logic, and multi-hop questions.
Anchor the model to sources and require it to admit uncertainty.
Only answer using information from the <sources></sources> block below. If the answer is not in the sources, respond exactly: 'Not found in provided sources.' Never fabricate.
The model hallucinates when it's forced to fill gaps. Give it permission to say 'I don't know'.
After answering, the model critiques its own answer as a hostile reviewer would.
Step 1: Answer the question. Step 2: Now play the role of a hostile fact-checker and list every claim in your answer that could be wrong. Step 3: Revise the answer.
The model catches its own errors when instructed to look for them — RLHF trained it to be agreeable, not accurate.
Show 2-5 input/output pairs. The model matches the pattern.
Example 1:
Input: {ex1_in}
Output: {ex1_out}
Example 2:
Input: {ex2_in}
Output: {ex2_out}
Now do this one:
Input: {real_input}
Output:Pattern matching from examples beats abstract rule descriptions for structured output tasks.
Specify exact format, length, and forbidden words up front.
Respond in exactly {n} bullet points. Each bullet: max 15 words. Use plain English. Never use the words: {forbidden_list}.Ambiguity in the ask produces ambiguity in the output. Constraint = clarity.
Force the model to present multiple perspectives and label them by source.
Present the {topic} debate as follows: (1) The mainstream consensus and its evidence. (2) The credible dissenting view and its evidence. (3) Fringe claims and why they lack support. Label each with named sources.AI models default to consensus. Explicit prompts for dissent surface information you'd otherwise miss.