Introduction
Prompt engineering is not static. What worked on GPT-3 does not work on Claude Opus or GPT-5.6. IBM ran 430,738 evaluations across eight popular prompting techniques on modern models. The results challenged almost every assumption the AI community had built up. Here is what the research actually says.
1. Never End Your Prompt With "Right?"
Cornell Tech tested 45 different AI models with one-word changes at the end of prompts. Both "right?" and "maybe?" bias the model toward your conclusion rather than letting it reason independently. The model does not have an opinion. It mirrors the direction your phrasing points.
Bad: I am deciding between renting and buying. Buying is the better choice, right?
Good: I am deciding between renting and buying. Compare both options for my situation.
2. Stop Using "Step by Step"
IBM tested the most popular prompting techniques, including the famous "Let's think step by step," widely recommended since 2022. It lost to asking normally. The winning combination was simply your question plus a brief role in a few words, no elaborate persona, no process instructions.
Bad: [Question]. Gather information, devise a plan, answer step by step.
Good: [Question] as a financial analyst.
3. Do Not Trust Confident Output
Microsoft's CTO audited 2.6 million references cited at the world's top AI conferences. One in four NeurIPS 2025 papers that passed expert peer review had a hallucinated citation, and reviewers scored those papers slightly higher. If trained academics cannot catch AI hallucinations in their own field, the average professional will not either. Every link, statistic, or citation needs to be verified at the source before it goes anywhere.
4. Keep Your Rules to Three Maximum
Meta tested 15 AI models on prompts with one to twelve rules. At eight rules, models followed all eight simultaneously only 5.7% of the time. Twelve of the fifteen models cannot reliably hold more than three rules at once. The solution is two runs, not lower expectations.
Bad first run: Write a LinkedIn post. Exactly 3 paragraphs. Under 150 words. No emojis. Include these three words. No competitor mentions. End with a question. Grade-6 reading level.
Good first run: Write a LinkedIn post. Include these three words. No emojis. End with a question.
Good second run: Check the draft against each requirement one at a time, then revise: 3 paragraphs, under 150 words, grade-6 reading level, no competitor mentions.
5. Goals Beat Examples
Researchers tested modern models with the practice of providing worked examples before the question. Mistral scored 74% with examples. When examples were removed and replaced with a clear goal, the score jumped to 83.8%. Modern models are constrained by examples, not helped by them.
Bad: You are a world-class strategist. Example 1. Example 2. Find a plan; answer step by step.
Good: My newsletter open rate fell from 40% to 28% over three months. The format, send time, and subject line did not change. What are the most likely causes? Ask me for any data you need before you answer.
6. Tell AI What Not to Do
Adding a short "do not" section to your prompt eliminates the filler, corporate-sounding language, and generic structures that make AI output instantly recognizable. Be specific about what to avoid, not just what to produce.
Bad: Write a summary of this report.
Good: Write a summary of this report. Do not use bullet points. Do not start with the word "This." Do not include recommendations, only findings.
7. Break Large Requests Into Sequences
Asking for a full report, presentation, or analysis in one prompt produces something shallow. Breaking the task into a sequence, first an outline, then each section, then a critique pass, produces something genuinely usable. Each prompt builds on the last and you stay in control at every stage.
8. Put the Most Important Instruction First or Last
Research consistently shows that models lose focus on information placed in the middle of long prompts. This is called the lost in the middle problem. The most critical constraint or goal should appear at the very beginning or very end of your prompt, never buried in the body where it gets underweighted.
Conclusion
Prompt engineering is not about tricks. It is about understanding how modern models process your input. Avoid leading questions, drop the step-by-step instruction, verify every fact, limit your rules to three, lead with a goal, tell AI what to avoid, break large tasks into sequences, and put what matters most at the start or end.
Subscribe to the Awesome Analytics newsletter for weekly AI insights.
