Finding the right grain size is important...
 |
| Mechanical sieves filter out grains at the correct sizes. |
... especially when you're trying to figure out how to write a prompt to answer your research question. Here are four aspects of crafting a well-working prompt.
When partnering with an AI for complex research tasks, the success of your inquiry often hinges on how you scope the problem. We hear a lot about the mechanics of "prompt engineering," but a far more vital skill is "abstraction engineering"—calibrating the exact altitude at which to fly your research question (RQ). Working effectively with a generative model requires finding the "Goldilocks zone" of detail: not too broad, not too narrow, but perfectly contextualized.
If you frame your task at too high a level of abstraction, the AI will hand you back a beautifully structured plate of platitudes. You’ll get the generic encyclopedia summary when you actually need a nuanced, critical analysis. Conversely, if you zoom in too far—dictating rigid micro-steps or demanding highly specific, obscure data points right out of the gate—you back the AI into a corner. When treated like a traditional relational database or forced to retrieve hyper-specific, unindexed numbers, the model is highly prone to hallucinating or failing outright.
The sweet spot lies in defining the intent and the boundaries of your research without over-constraining the AI's ability to synthesize. You want to give it enough conceptual context to act as an intelligent thought partner, while setting clear parameters to keep it anchored to verifiable reality.
Let’s look at a concrete example. Imagine you’re investigating the historical impact of extreme weather on California's coastline.
Too Abstract: A prompt like: [Tell me about coastal erosion in California] is a bit too open-ended. The AI will generate a high-level, generic overview summarizing basic geological concepts and mentioning climate change. It’s structurally sound, maybe even more-or-less correct, but practically useless for serious research.
Too Granular: Consider this prompt by contrast: [What was the exact volume of sand, in cubic yards, lost from the southern end of Half Moon Bay, California between November 12 and November 15, 1983?] Here, you’re asking a generative text model to act as a raw data repository and data analyst. It will likely confidently invent a plausible-sounding number, immediately leading your research astray. This is a great way to generate junk quickly.
The Right Level of Abstraction: A much better prompt: [As a coastal geologist, I am researching the impact of the 1982-1983 El Niño storms on coastal erosion in Northern California. Can you synthesize the major geological impacts on the coastline during that specific winter, and then suggest which state agencies, archives, or specific scientific databases I should query to find the raw historical wave-height and sand-loss data for Half Moon Bay?]
This final approach is scoped perfectly. It leverages the AI for what it does best—synthesizing complex historical events and mapping the conceptual landscape—while strategically recognizing its limitations. By asking the AI to point you toward the right primary sources rather than demanding it be the primary source, you are utilizing it as an expert research librarian. This accelerates your workflow without compromising the integrity of your methodology.
2. Communicate your intent clearly enough for reliable hand-off to an AI.
We often treat AI like a mind reader. It is not. It’s more like a wildly enthusiastic, highly literal intern who just drank six espressos and wants to please you immediately. The critical moment in any AI-assisted research task isn't the underlying algorithm; it's the point where you transfer your beautifully complex, nuanced research goal from your brain into a text box. If you don't communicate your intent clearly, the AI will happily sprint off in the wrong direction and return milliseconds later with a pile of beautifully formatted, profoundly unhelpful text.
To successfully hand off a task, you have to explain the why alongside the what and maybe add a dash of context about who you are and what you expect.
The AI lacks the implicit context of your day-to-day life. It doesn't know you’ve spent three weeks agonizing over a methodology, nor does it know if you are writing a rigorous literature review or simply trying to settle a bar bet.
Here’s a remarkably common failure mode. Suppose you’re researching the history of urban sanitation (a riveting topic, I know) and you have a stack of primary source documents about 19th-century London. Here’s how to approach it:
Too Abstract: If your hand-off prompt is simply, [Summarize these papers], the AI will cheerfully oblige. You’ll get a perfectly bland, high-school-level essay about cholera, bad smells, and the River Thames. It’s historically accurate, but completely useless for actual research.
Too Granular: You might overcorrect and try: [Extract every mention of 'sewer pipe diameter' from these texts.] Now you have a sterile list of numbers completely divorced from their historical context. Also useless.
A reliable hand-off requires stating your overarching intent so the AI knows exactly what kind of intellectual heavy lifting it needs to do.
The Right Level of Abstraction: A better approach looks like this: [I am writing an academic paper comparing the municipal funding models of 19th-century London and Paris. I am specifically interested in how they paid for public works. Read these papers on London's sanitation system and extract only the sections detailing the financial instruments, bond issuances, or tax levies used to fund the Bazalgette sewer network. Explicitly ignore the medical history of cholera outbreaks.]
Notice the difference? You’ve given the AI a job description, a specific destination, and a "do not enter" sign for the irrelevant stuff. By explicitly defining your intent—what you are doing and why you are doing it—you constrain the model's infinite possibilities into a highly targeted research instrument. Treat the AI like a brilliant but amnesiac colleague who just walked into the middle of your project meeting. Tell them exactly what the end goal is before you put them to work, and you might actually get a useful result.
3. Evaluate the results you get back from your AI
If there is one universal law of generative AI, it is this: it will confidently hand you a fabricated answer with the serene, unshakeable certainty of a mediocre undergraduate who just skimmed a Wikipedia summary or handed in an AI generated output without reading.
This becomes a massive problem in serious research because, out here in the real world, "ground truth" isn't a magical, pristine spreadsheet handed down by a benevolent universe. Real data is often incomplete, contradictory, or buried under a mountain of historical noise.
So, how do you trust an AI assistant when you don't have the perfect answer key to check its work?
You have to build-in your evaluation strategy, thinking about it before you hit "submit," shifting your focus from verifying a single final answer to stress-testing the model's methodology.
Let’s say you are trying to piece together the economic history of a regional industry—for instance, the apple export market in 1920s Washington state. The historical records are a disaster. Farm manifests were lost in fires, different counties used different metrics (bushels versus crates versus trainloads), and local agricultural boards routinely exaggerated their yields to look good. The ground truth is inherently noisy.
Too Abstract: Here is what the "too abstract" version of our 1920s Washington apple research looks like:
[Summarize the state of the Washington apple export economy in the 1920s and tell me how successful it was.]
When faced with a prompt this broad, the AI will happily oblige by synthesizing a beautifully written, highly readable narrative. It will tell you about the booming agricultural sector, the arrival of new rail lines, and the indomitable spirit of the Pacific Northwest farmer. It might even throw in a generic quote about the crispness of a Red Delicious. It will sound incredibly authoritative, like a velvet-voiced narrator on a PBS documentary.
And it will be completely useless to you as a researcher.
Why? Because at this level of abstraction, the AI actively hides the noisy ground truth from you. Instead of dealing with the messy reality that Chelan County measured their yield in "crates" while Yakima County measured in "freight cars" and half the records burned down in 1926, the model simply smooths all that chaotic data into a neat, frictionless trendline.
And asking for a judgement call (“tell me how successful it was”) is a beginner’s error. The AI doesn’t have a point-of-view, but it’ll make one up.
All of this papers over the contradictions and missing farm manifests with plausible-sounding historical clichés. Because you asked a vague question, you get a generalized synthesis, leaving you with absolutely no way to evaluate the accuracy of its claims. You can't audit the model's work because the AI has abstracted away all the actual evidence. You asked for a rigorous economic history, and it handed you a tourism brochure.
Too Granular: If you approach this at the wrong level of detail and ask the AI, [What was the exact total tonnage of apples exported from Washington in 1924?] the model will gladly average out the historical lies, hallucinate a plausible-sounding integer, and present it as absolute fact. Because the underlying data is a mess, you have no way to evaluate whether the AI's number is a brilliant synthesis or a total fabrication. You are trapped in a dead end of misplaced trust.
The Right Level of Abstraction: A better approach—scoping the task to account for that noisy reality—looks like this: [I am researching 1920s apple exports in Washington state, but the historical county records are contradictory and use mixed units. Here is a text dump of five different agricultural reports from 1924. Please extract the export claims from each, standardize the units into tons where possible, and explicitly flag any mathematical discrepancies—for example, if a county claims to have exported more apples than they had arable acreage to grow them. Do not attempt to give me one final definitive number; just map the contradictions. Please include all references to source materials.]
Notice the shift. You haven't asked the AI to find the "truth," because the truth is currently unknowable. (This means that prompts like “tell me just the facts” are fundamentally hopeless.) Instead, you've asked it to structure the results with a little fact-checking. With this approach, you can actually evaluate the AI's output reliably: did it catch the logical discrepancy between acreage and yield? Did it convert the units correctly? By adjusting the level of your RQ, you transform the AI from a highly suspect oracle into a tireless research assistant helping you audit a messy reality.
Also, asking for the references is an important step. (Be sure to check that they’re real!)
4. Plan to iterate on your prompt.
Just as in the old days of Google search (meaning, last year), there is a persistent, romantic myth in the world of generative AI that the "perfect prompt" exists. We want to believe that if we just arrange our words with the exact right term choice and a bit of alchemy, the AI will do the magic and hand down a flawless, publication-ready analysis on the first try.
Let me disabuse you of that notion right now: your first prompt is almost always going to be wrong or just slightly misaligned with the AI. And that’s entirely okay.
Working with AI isn’t a vending machine transaction; it’s a conversation. Plan to iterate.
It is exceedingly rare that your initial text string will capture the full, nuanced intent of your research goal. You are going to find subtle misinterpretations, bizarre blind spots, and moments where the model took your slightly ambiguous phrasing and ran off in the wrong direction.
The real skill is treating that first output not as a final answer, but as a diagnostic tool to figure out how to calibrate the exact level of detail your RQ actually requires.
Here’s how this iterative process helps you find the right level of abstraction for your RQ. Suppose you’re researching the engineering failures of the 1858 transatlantic telegraph cable.
Too Abstract: [Why did the 1858 telegraph cable fail?] The AI gives you a decent, if sleepy, overview about a guy named Wildman Whitehouse applying too much voltage. It’s too abstract and open-ended. But reading it, you realize you actually want to know about the physical degradation of the cable itself. (And it IS a great way to come up with topics for additional deep-dive research.)
Too Granular: [What was the exact chemical breakdown rate of the gutta-percha insulation on the 1858 cable on August 15th?] Now you’ve zoomed in too far. The AI panics at the hyper-granularity, either hallucinating a fake chemical decomposition story or apologizing that it doesn't have the daily logs. But now you know your boundaries.
The Right Level of Abstraction: Here’s a much better, more appropriate level of detail for your prompt: [For a paper on the history of telegraphy to be submitted to the local newspaper I am researching the physical degradation of the 1858 transatlantic cable. My previous searches indicated that high voltage destroyed the gutta-percha insulation. Can you synthesize the historical consensus on how the seawater interacted with the compromised insulation to cause the final short circuit? After your summary, please list three historical archives or electrical engineering journals where I could find primary source correspondence about the cable's physical testing]
By iterating, you’ve discovered the sweet spot. You provided the target audience (local newspaper), the context (historical consensus on insulation failure), asked for a synthesis appropriate for a language model, and then directed it to help you find the primary sources for the hyper-granular data. You didn't fail on your first prompt; you simply used it to map the territory so you could finally ask the right question.
Keep searching!
No comments:
Post a Comment