A lot of analysts think ChatGPT got worse this year. The model got stronger. The answers got weaker. That sounds impossible. It is not. And once you see why, you will prompt for research completely differently.
Here is what actually changed. Two things, and both matter for the work you do. First, there is now an invisible router sitting in front of the model. You type a question, you hit enter, and something decides which engine answers you before the model ever does. The reporting on this is blunt. That router often picks the cheapest, fastest, weakest option. So a high-stakes question can quietly get a low-effort answer and you would never know. Second, the model now follows instructions literally. The old versions guessed what you meant when your prompt was lazy. This one stopped guessing. Put those two together and you have the whole story. A vague prompt gets routed to a weak engine, and then taken literally. That is why loose prompting feels worse now. So let me walk you through the research prompt that fixes it, one layer at a time.
Start with the question itself. "Analyze this stock" is the vague prompt. It always was. The old models hid that from you by guessing what you wanted. This one will not. So layer one is to force the model to actually think. There is a simple trigger that does it. You add a phrase like "think hard about this." That nudges the router toward a real reasoning engine instead of a snap answer. And reasoning matters for one specific reason. Deeper reasoning surfaces second-order effects you had not considered. For research, that is the entire point. The first-order answer is the obvious one. The edge is in the second-order one. So for any high-stakes question, make the model reason. Do not let it answer fast.
Layer two is the analyst standard, and it is non-negotiable. Tell the model exactly what it is reading. Name the source. Name the date. Name the exact filing section. The latest ten-Q. The second-quarter revenue note. The risk factors. The management discussion and analysis. If the model is summarizing a filing, it should tell you which part of the filing. And here is the rule underneath it. If it cannot show you the source, it is not research. It is a guess in a confident voice. A literal model is actually a gift here, because it will hold to the source you give it. So give it a real one.
Layer three. Stop dumping everything into one paragraph. The model now reads structure literally, so use structure on purpose. Think of it as labeled boxes. This box is context. This box is the task. This box is the output format. You wrap each one in its own tag so the model cannot blur them together. Background goes here. The question goes here. The format you want the answer in goes here. When the boxes are labeled, comprehension jumps and the answer gets sharper. This is worth saving as a template, because you reuse the same boxes every single time. Build it once. Paste it on every name you cover.
Layer four is the one most people skip, and for markets it is the most important. Tell the model to separate fact from inference from opinion. Every line in the output gets a label. This is a fact from the filing. This is the model's inference. This is opinion. And any number it cannot tie to a source, it has to flag. Why does that matter so much? Because a polished paragraph hides the difference between what is true and what the model assumed. The split forces it into the open. Now you can see exactly where the judgment calls are. And those judgment calls are your job. Not the model's.
Last layer. The model is good at one more thing. Critiquing its own work. So make it do that before you ever see the draft. You tell it upfront to build its own rubric for what a world-class research brief looks like. Then grade its first draft against that rubric. If it scores a six, it rewrites. A seven, it rewrites again. It keeps going internally until the draft holds up. You are not accepting the first answer and fixing it by hand. You are making the model iterate to a high bar before it hands you anything. And you close every prompt with one line. List what to verify next. That single line keeps you honest. The brief is a starting point, not a verdict.
Let me make it concrete with two examples. Example one. You are reviewing a quarterly filing. The bad prompt is "summarize this earnings report." The research prompt sets the role, names the filing and the date, tells it to think hard, pulls the deltas against the prior quarter on revenue drivers, margins, guidance, and risk language, labels each line fact, inference, or opinion, and ends with what to verify next. Same model. Completely different output. Example two. You are building an overview of a sector. The bad prompt is "tell me about the enterprise AI market." The research prompt asks it to build a rubric for a world-class market overview first, draft against named sources and dates, iterate until it scores top marks, then hand you a list of open questions. Notice what neither example does. Neither one tells you to buy or sell anything. They build the brief. You make the call.
So here is the whole thing in one breath. The model got literal and the router got cheap. That broke lazy prompting and rewarded disciplined prompting. Raise the stakes. Name the source. Label the boxes. Force the split. Make it grade itself. And these stack. You can run all five in a single prompt. The model constructs the brief faster and cleaner than you could. But the assumptions, the disconfirming evidence, and the decision are still yours. That part never moved.
If you want the research prompt template with all five layers built in, comment the word RESEARCH and I will send it over. And if you want one AI research workflow like this every week, subscribe. Educational content only. Not financial, legal, tax, or investment advice. Examples are workflow illustrations, not recommendations.