Most prompting advice was never tested on anything. These 4 rules were, on real models, by people who published the numbers. Each one comes with the study, a before and after, and a prompt you can copy.
The tip everyone shared in 2023 was "take a deep breath and think step by step." It was tested, and it worked. On a model nobody uses anymore.
That is the problem with most prompting advice. It was tested once, on one model, and then repeated for 3 years while the models changed every few months. Or it was never tested at all.
This page keeps only the rules with a study behind them, run on the models you use today, with the numbers published. There are 4. Each one has the study, a before and after, and a prompt you can copy.
Every card has the study at the top, a before and after underneath, and a prompt you can copy. Start with the rule you are breaking today. Most people are breaking number 2.
A Cornell Tech researcher tested 45 models with one-word changes at the end of a question. "[X] is the better choice, right?" and "[X] is the better choice, maybe?" both pulled the models toward agreeing with you. The lean in your question decides the answer more than the model's own judgment does.
The fix. Ask it the way you would ask a friend you want the truth from. Neutral, no yes or no, no lean. If you already have a favorite, ask what is wrong with it.
I'm deciding [the decision, e.g. whether to raise my prices in September]. Compare [option A] and [option B] for my situation: [2 or 3 facts about your business that matter here]. Then tell me what is wrong with [the option I'm leaning toward]. Do not tell me which one I want to hear.
Meta tested 15 models (GPT, Claude, Gemini and 12 others) on prompts with 1 to 12 rules. At 8 rules, a model gets each rule right 41% of the time. It gets all 8 right at once 5.7% of the time. 12 of the 15 models could not hold more than 3.
The fix. Pick your 3. Put the rest in a second message and make it check the draft against them one at a time. Same rules, 2 runs, and it holds all of them.
First message: Write [the thing, e.g. an Instagram caption for my new offer]. Must include: [the 3 things that matter most, e.g. the price, the date, a question at the end]. Second message, after the draft comes back: Check your draft against each of these one at a time, then revise: [the rest of your rules, e.g. under 120 words, no emojis, plain English, no exclamation marks].
IBM ran 430,738 evaluations across 8 prompting techniques. "Let's think step by step," the trick that made prompting famous, lost to asking normally. The winner was the question plus your role in a few words. Newer models already reason before they answer, so the instruction just gets in the way.
The fix. Ask the question. Add a role in a few words. Delete "gather information, devise a plan, answer step by step" and everything like it.
[Your question, asked plainly] ... as a [the role that fits the job, e.g. bookkeeper for a business of one, or wedding photographer with 10 years of experience].
Researchers ran a Mistral model with the standard prompt full of worked examples and scored 74%. They deleted the examples and scored 83.8%. Examples teach the model to copy the shape of the example. The goal teaches it what you want back.
The fix. Say what is true, say what you want, and let it ask for what it is missing. No "you are a world class strategist," no pasted examples.
[The situation in 2 sentences, with the real numbers, e.g. My newsletter open rate fell from 42% to 31% over 3 months. I send one issue every Tuesday and changed nothing about the format, the send time or the subject line.]
What are the most likely causes?
Ask me for the data you need before you answer.
There is a 5th rule from the same pile of research, and it is the one that costs the most when you skip it. Do not trust confidence. A Microsoft audit of 2.6 million references at a top AI research conference found 1 in 4 papers had a citation that did not exist. Those papers had passed expert peer review. The reviewers missed it and scored them slightly higher.
If professional reviewers cannot catch a made-up source, you will not either by reading. So open every link, every time. If a number matters, ask where it came from and go look.
This template folds all 5 into one prompt. Paste it once and fill in the brackets.
[Your question, asked plainly, with no lean and no yes or no] Answer as a [role in a few words]. The goal: [what should be true when this is done]. What you're working with: [2 or 3 real facts or numbers]. Rules, 3 at most: [rule 1], [rule 2], [rule 3]. Ask me for anything you need before you answer. Mark anything you are not sure of, and give me the source for any number or fact so I can check it. After the draft, I'll send the rest of my rules and you'll check the draft against them one at a time.
The second message is always the same 2 lines: "Check your draft against each of these one at a time, then revise:" and your remaining rules.