Using LLMs as Research Assistants

I need to tell you about the first time I used an LLM for research. I had a question about the interaction between two regulatory frameworks, and instead of spending an hour reading primary sources, I asked the model. It gave me a confident, well-structured, beautifully articulated answer. It was also wrong in two significant ways — one a subtle mischaracterization of a policy, the other an outright fabrication of a court case that did not exist. The answer read so well that if I had not happened to know one of the areas reasonably well, I would have believed it entirely and built further conclusions on a foundation of articulate nonsense.

This is the fundamental tension of using LLMs for research. They are extraordinarily useful. They are also unreliable in ways that are particularly dangerous because their failures look exactly like their successes. A wrong answer from a search engine is usually obviously wrong — a broken link, a clearly irrelevant result. A wrong answer from an LLM is fluent, confident, and formatted exactly like a right answer.

So let me be direct: this chapter is about how to use LLMs as research tools effectively, and “effectively” means with a clear-eyed understanding of both their capabilities and their failure modes. If you use them as I am about to describe, they will make you a significantly better researcher. If you use them as oracles that produce truth on demand, they will make you confidently wrong.

The Right Mental Model

Stop thinking of an LLM as a search engine. Stop thinking of it as an encyclopedia. Start thinking of it as a very well-read research assistant who is eager to help, occasionally makes things up, and will never tell you when they are unsure.

This mental model is useful because it correctly calibrates your expectations. A good research assistant can:

• Help you brainstorm angles you had not considered
• Produce rough first drafts that you then verify and refine
• Synthesize information from multiple domains you may not be equally familiar with
• Explain complex concepts in accessible language
• Generate reading lists and suggest search terms
• Play devil’s advocate when you need to stress-test an idea

A good research assistant cannot:

• Be your sole source for factual claims
• Replace reading primary sources
• Guarantee that any specific detail is accurate
• Know what they do not know
• Tell you when they are guessing vs. when they are confident for good reasons

If you treat the LLM as the former and not the latter, you will be fine. If you blur the line, you will eventually publish, present, or act on something that is not true.

Prompting Strategies That Actually Work

The difference between a mediocre LLM research session and a great one is almost entirely in how you prompt. Most people prompt LLMs the way they use search engines — short queries expecting direct answers. This is like hiring a research assistant and then only ever sending them one-line emails. You get back something, but it is not what they are capable of.

Here are the prompting strategies I use most frequently, with examples.

Strategy 1: The Landscape Survey

When you are starting research on a new topic, you do not need depth yet. You need the shape of the field: who the main players are, what the key debates are, where the boundaries lie.

Prompt:

I'm starting to research [topic]. I need a landscape survey, not a deep
dive. Give me:

1. The 3-5 most important subtopics or questions within this area
2. The main schools of thought or competing perspectives
3. Key terms I should know to search effectively
4. Where the field is currently unsettled or actively debated
5. The most commonly cited foundational works (books, papers, or reports)

Be specific. If you're uncertain about something, say so rather than
guessing.

This prompt does several things right. It defines the output format explicitly. It asks for structure rather than a single narrative (which discourages the LLM from constructing a coherent-sounding story that papers over genuine complexity). And it explicitly invites uncertainty, which LLMs will almost never volunteer unless asked.

Strategy 2: The Steelman

You have a position on something. Maybe it is a business decision, a technical choice, or a policy opinion. You want to stress-test it.

Prompt:

I believe [your position]. I want you to steelman the strongest possible
counterargument. Don't give me a weak or straw-man version of the
opposing view. Give me the version that would be articulated by the
smartest, most informed person who genuinely disagrees with me.

Specifically:
- What evidence would they cite?
- What assumptions in my position would they challenge?
- What are they seeing that I might be missing?
- Where is my position most vulnerable?

This is one of the highest-value uses of an LLM. Not because the model has better judgment than you, but because it can access a broader range of perspectives faster than you can by reading. The steelman prompt forces the model to construct the best opposing case rather than a balanced “on one hand, on the other hand” summary that does not really challenge anything.

Strategy 3: The Blind Spot Check

Related to the steelman, but broader. Use this when you are in the middle of research and want to make sure you have not missed an entire dimension.

Prompt:

I'm researching [topic]. So far, I've been focusing on [aspects you've
covered]. What am I likely missing? Give me five perspectives or angles
on this topic that I probably haven't considered, especially from:

- Adjacent fields or disciplines
- Different geographic or cultural contexts
- Historical precedents
- Practical implementation concerns
- Ethical or second-order consequences

The explicit list of categories is important. Without it, the model tends to give you five variations on the same perspective. By naming the categories, you force diversity in the output.

Strategy 4: The Reading List Generator

LLMs can generate reading lists, but naive approaches produce a mix of real and fabricated sources. Here is how to get useful results:

Prompt:

I need a reading list on [topic], focused on [specific aspect]. Give me
10-15 recommendations in these categories:

- 2-3 foundational/classic works that anyone studying this topic should know
- 3-4 recent (last 3 years) publications that represent the current state
  of thinking
- 2-3 accessible introductions for someone coming from [your background]
- 2-3 contrarian or minority-view works that challenge mainstream thinking

For each, give me: title, author, year, and a one-sentence description of
why it's specifically relevant.

IMPORTANT: Only include works you are confident actually exist. If you are
not sure whether a specific work exists, describe the type of work I
should look for instead of fabricating a specific title.

That last paragraph is critical. LLMs will confidently generate plausible-sounding titles and authors that do not exist. The explicit instruction to flag uncertainty does not eliminate this problem, but it reduces it meaningfully. You should still verify every recommendation against a library catalog or search engine before investing time in hunting it down.

Strategy 5: The Explainer

When you need to quickly understand a concept from an unfamiliar domain, LLMs excel at calibrated explanations.

Prompt:

Explain [concept] to me. I have a strong background in [your field] but
limited knowledge of [the concept's field]. Use analogies from my field
where possible. Cover:

1. What it is, in one paragraph
2. Why it matters, in the context of [what you're researching]
3. The most common misconceptions about it
4. How experts in this area actually use/apply it
5. Where the simple explanation breaks down and the reality is more
   complex

Point 5 is doing the heavy lifting here. Without it, you get a clean explanation that makes the concept seem simpler than it is. With it, you get an explanation that tells you where the simplification stops being reliable — which is exactly what you need to know when you are using this understanding to make decisions.

Strategy 6: The Pre-Mortem

Before making a decision based on your research, use this:

Prompt:

I'm about to [decision/action] based on the following reasoning:
[your reasoning].

Conduct a pre-mortem: assume this decision turns out to be wrong or
produces bad outcomes. What are the three most likely reasons it failed?
For each:

- What did I get wrong or overlook?
- What changed that I didn't anticipate?
- What information would have led me to a different decision?

This is valuable not because the LLM can predict the future, but because it can generate plausible failure scenarios that you might not think of when you are committed to a course of action. It is a structured way to break out of the tunnel vision that naturally develops during research.

Structuring a Research Session

Random prompting produces random results. A structured research session produces useful ones. Here is the workflow I use:

Phase 1: Orientation (10-15 minutes)

Start with the Landscape Survey prompt. Read the output critically — not to learn facts, but to get a map of the territory. Note the terms, names, and frameworks mentioned. Do not assume any of them are real yet.

Follow up with 2-3 clarifying questions based on the landscape survey. “You mentioned [concept] — can you explain how that differs from [related concept]?” “You listed [person] as a key thinker — what is their main contribution?”

Phase 2: Exploration (20-30 minutes)

Now go deeper on the specific aspects that are most relevant to your needs. Use the Explainer prompt for unfamiliar concepts. Use the Blind Spot Check to make sure you are not missing important angles.

This is where multi-turn conversation becomes important. Do not start a new conversation for each question — build on the context you have established. The model’s responses will be more coherent and useful when it has the full context of your research session.

A sample exchange:

You: You mentioned that there are three main approaches to [topic]. Let's
go deeper on the second one. What are its specific strengths and
weaknesses compared to the others?

LLM: [response]

You: You said the main weakness is [X]. Can you give me a concrete example
of this weakness manifesting in practice? And who has written most
critically about this limitation?

LLM: [response]

You: That example is helpful. Now steelman the second approach — if
someone were defending it against that criticism, what would their best
argument be?

Notice the pattern: each prompt builds on the previous response, adding specificity and pushing for nuance. This is much more effective than a series of disconnected questions.

Phase 3: Stress-Testing (15-20 minutes)

Now use the adversarial prompts — Steelman, Pre-Mortem, and the Blind Spot Check. By this point, you have probably started forming opinions about your topic. This is the phase where you deliberately try to break those opinions.

Phase 4: Verification and Reading List (10-15 minutes)

End the session by generating a reading list for primary source verification. Use the Reading List Generator prompt. Also ask:

Based on our conversation, what are the three most important factual claims
I should independently verify before relying on this research? For each,
suggest where I would find authoritative primary sources.

This creates your verification checklist. The LLM has given you a first-draft understanding; now you need to confirm the key facts through sources you can actually trust.

Total session time: 55-80 minutes

This produces a landscape understanding, a nuanced view of the key issues, a set of stress-tested conclusions, and a verification checklist. Compare that to 55-80 minutes of undirected reading and you will see why this approach is powerful.

The Verification Problem

I keep returning to verification because it is the single most important practice in LLM-assisted research, and it is the one most people skip.

LLMs have several failure modes that are relevant to research:

Hallucination: The model generates plausible-sounding information that is entirely fabricated. This includes fake citations, non-existent studies, invented statistics, and fictional quotations attributed to real people. The frequency of hallucination varies by model and topic, but it is never zero.

Outdated information: Models have training data cutoffs. They may present outdated information as current, or miss recent developments that change the picture significantly.

Majority bias: Models tend to reflect the most common perspective in their training data. Minority viewpoints, emerging research, and contrarian positions may be underrepresented or presented as more fringe than they actually are.

Confident uncertainty: Models almost never say “I don’t know” unless you explicitly create space for them to do so. When they are uncertain, they typically generate the most probable-seeming answer and present it with the same confidence as a well-established fact.

Coherence over accuracy: Models are optimized to produce coherent, well-structured text. When accuracy and coherence conflict — when the truth is messy, contradictory, or uncertain — the model will often choose the cleaner narrative.

Given these failure modes, here is a practical verification approach:

Verify all specific claims. If the LLM says “a 2019 study by [author] found [result],” look up the study. Confirm the author, the year, the result, and the context. This takes 2-3 minutes per claim and has saved me from citing non-existent research more times than I am comfortable admitting.

Cross-reference with traditional search. After your LLM session, take the key conclusions and search for them using a regular search engine. Look for confirming AND contradicting evidence. If you cannot find independent confirmation for a key claim, treat it as unverified.

Check the edges. LLMs are most reliable on well-established, mainstream topics. They become less reliable on recent events, niche topics, topics where the evidence is genuinely contested, and anything involving specific numbers, dates, or quotations. Apply extra scrutiny in these areas.

Watch for suspiciously clean narratives. Real research topics are messy. If the LLM gives you a neat, tidy story with no contradictions, caveats, or loose ends, be suspicious. Reality is rarely that clean. Prompt for complexity: “What parts of this are more uncertain or contested than your summary suggested?”

When to Use an LLM vs. Traditional Search

LLMs and search engines are complementary tools, not substitutes. Here is when each excels:

Use an LLM when:

• You need a synthesis across multiple sources or domains
• You are exploring a new topic and need orientation
• You want to brainstorm perspectives, angles, or hypotheses
• You need an explanation calibrated to your level of expertise
• You want to stress-test your own thinking
• You need to quickly understand how concepts from different fields relate

Use traditional search when:

• You need specific, verifiable facts (dates, statistics, quotations)
• You need the most current information on a topic
• You need primary sources (the actual paper, the actual regulation, the actual data)
• You need to confirm that something the LLM told you is true
• You are looking for a specific document you know exists
• You need information about recent events

Use both when:

• You are conducting serious research on any topic. Start with the LLM for orientation and synthesis, then use search for verification and primary sources.

The workflow looks like this:

LLM → Orientation and synthesis → Key claims identified
         ↓
Search → Verify claims → Find primary sources → Fill gaps
         ↓
LLM → Refine understanding based on verified information
         ↓
Search → Final verification pass

This loop can run two or three times in a research session, with each pass producing a more accurate and nuanced understanding.

Combining LLMs With Other Tools

LLMs are most powerful when combined with other research tools rather than used in isolation. Here are some combinations I use regularly:

LLM + Google Scholar

Use the LLM to identify key concepts and search terms, then use Google Scholar to find actual papers. The LLM-generated reading list gives you authors and concepts to search for; Scholar gives you the real papers and their citation networks. Follow the citations — if a paper is heavily cited, it is probably worth reading regardless of whether the LLM mentioned it.

LLM + Domain-Specific Databases

Every field has its specialized databases. PubMed for biomedical research. SSRN for social science working papers. arXiv for physics and math. IEEE Xplore for engineering. Use the LLM to understand what you should be searching for, then use the domain database to find it. The LLM excels at translating between your terminology and the field’s terminology, which makes your searches in specialized databases much more effective.

LLM + RSS/News Aggregation

When you encounter a topic in your daily information triage (Chapter 9), use the LLM to quickly get background context. “I just saw a headline about [topic]. Give me a two-paragraph explanation of why this matters and what the key context is.” This turns a headline into an informed assessment in about 30 seconds.

LLM + Collaborative Tools

If you are researching in a team, use the LLM to generate a structured briefing document that you can share and collaboratively refine. The initial LLM output becomes a draft that the team annotates with their own knowledge, corrections, and additional sources. This is much more efficient than having each person do independent research and then trying to synthesize.

LLM + Your Own Notes

This is underrated. If you maintain research notes, a knowledge base, or a personal wiki, you can paste relevant sections into the conversation as context. “Here are my notes on [topic] from the last six months. Based on this, what gaps do you see in my understanding? What recent developments should I update my notes with?”

Multi-Turn Conversation Strategies

Single-prompt interactions with LLMs are like asking a question at a conference Q&A. You get an answer, but it is generic and uncalibrated. Multi-turn conversations are like sitting down with the speaker for coffee afterward. The quality goes up dramatically.

Here are specific strategies for productive multi-turn research conversations:

Build context progressively. Start broad, then narrow. Each turn should add specificity to what you are asking. “Tell me about X” → “Interesting, let’s focus on the [specific aspect] you mentioned” → “How does that interact with [Y, which you already know about]?”

Challenge, do not just accept. When the model gives you an answer, push back on it. “You said [X]. But doesn’t that conflict with [Y]? How do you reconcile those?” This forces the model to engage with complexity rather than defaulting to simple narratives.

Introduce your own knowledge. The model does not know what you already know. When it tells you something you are familiar with, say so: “I know about [X] already — I’m specifically interested in [more specific aspect].” This prevents the model from spending time on basics and pushes it toward the territory where it can actually add value.

Redirect when the model goes off track. LLMs sometimes drift toward tangential topics, especially in long conversations. It is your job to steer: “That’s interesting but not what I need right now. Let’s get back to [the thing you actually care about].”

Summarize periodically. In long conversations, ask the model to summarize the key conclusions so far. This serves two purposes: it gives you a checkpoint to verify that the model’s understanding matches yours, and it refreshes the model’s context window so that earlier important points are not lost.

Here is an example of a productive multi-turn research conversation about a topic I was recently exploring — the effectiveness of different approaches to reducing misinformation:

Me: I'm researching approaches to reducing misinformation on social
platforms. Give me a landscape survey of the main strategies that have
been proposed or implemented.

LLM: [Provides overview of fact-checking, algorithmic downranking,
media literacy programs, prebunking, community notes models, etc.]

Me: You mentioned prebunking. I'm less familiar with this than the
others. Explain the research basis for it and where experts disagree
about its effectiveness.

LLM: [Explains prebunking, cites inoculation theory, discusses
effectiveness debates]

Me: Interesting. I've been assuming that algorithmic solutions are
more scalable than educational ones. Steelman the opposite view —
that educational approaches like prebunking are actually more
effective at scale than algorithmic interventions.

LLM: [Constructs strong case for educational approaches, including
arguments about algorithmic approaches being adversarially brittle,
creating censorship concerns, and not addressing the root vulnerability]

Me: Good points. Now, what am I missing entirely? What approaches to
this problem am I not even considering?

LLM: [Introduces supply-side interventions, economic incentive
redesigns, platform interoperability as an approach, and structural
media ecosystem changes]

Me: The economic incentive angle is new to me. Let's go deeper there.
Who are the key thinkers, and what specifically are they proposing?

LLM: [Goes deeper on economic approaches]

Me: Based on everything we've discussed, what are the three claims
I should verify independently, and where would I find authoritative
sources for each?

Notice how each turn builds on the previous one, gradually moving from breadth to depth while periodically checking for blind spots. This conversation covered more ground in 20 minutes than an hour of undirected reading would have.

Common Mistakes to Avoid

I have made most of these mistakes personally, some of them repeatedly. Sharing them here so you can make different ones.

Mistake 1: Treating LLM output as final. Never use LLM-generated text as your finished product without verification and revision. The model’s output is a first draft. A good first draft, often, but a first draft. If you are writing a report, making a presentation, or publishing anything based on LLM-assisted research, the verification step is not optional.

Mistake 2: Not providing enough context. The more the model knows about your specific situation, background, and needs, the more useful its output will be. “Tell me about quantum computing” produces a generic overview. “I’m a software engineer evaluating whether my company should invest in quantum-resistant cryptography. What do I need to understand about the current state of quantum computing, specifically the timeline for threats to current encryption?” produces something you can actually use.

Mistake 3: Asking leading questions. If you ask “Isn’t it true that [your existing belief]?”, the model will almost always agree. This is confirmation bias with extra steps. Ask open questions. Ask for counterarguments. Ask what the model would say if you were wrong.

Mistake 4: Ignoring the model’s limitations on recency. If your research involves anything that has changed in the last year or two, the model may be working with outdated information. Always ask: “What is your training data cutoff? Is there anything about [topic] that has likely changed since then?” And then verify current status through search.

Mistake 5: Using one long session instead of multiple shorter ones. In very long conversations, LLMs can develop what I think of as “conversational drift” — they start subtly adjusting their perspective to match what they perceive as your expectations. Break major research projects into multiple sessions. Start fresh conversations when you change subtopics. Compare outputs from different sessions to check for consistency.

Mistake 6: Not saving your sessions. LLM conversations are ephemeral by default. If you conduct a good research session, save the transcript. You may want to revisit it, share it with colleagues, or use it as context for a future session. Most LLM interfaces let you export conversations or at least copy them. Do this before you close the window.

A Template for Your First Session

If you have not used LLMs for research before, or if your previous attempts were unsatisfying, try this template for your next research question:

Turn 1: Landscape Survey (the prompt from Strategy 1 above)

Turn 2: Pick the most relevant subtopic from the landscape survey
        and ask for a deeper explanation using Strategy 5 (The Explainer)

Turn 3: Use Strategy 3 (Blind Spot Check) — what are you missing?

Turn 4: Dive deeper on whatever the blind spot check surfaces

Turn 5: Use Strategy 2 (Steelman) on whatever conclusion you are
        forming

Turn 6: Use Strategy 4 (Reading List Generator) to get primary sources

Turn 7: Ask for a verification checklist — what should you confirm
        independently?

Seven turns. Thirty to forty-five minutes. You will end up with a structured understanding of the topic, a list of challenges to your initial thinking, a reading list for deeper investigation, and a clear set of claims to verify. This is what a good research assistant produces — not answers, but a dramatically accelerated path to answers.

The Bottom Line

LLMs are not a shortcut to knowledge. They are a shortcut to a structured first draft of understanding, which you then refine through verification, primary sources, and your own critical thinking. Used this way, they are genuinely transformative tools for research. Used as substitutes for actual thinking, they are fluent bullshit generators.

The prompting strategies in this chapter work because they are designed to exploit what LLMs are good at (synthesis, brainstorming, generating diverse perspectives, explaining across domains) while compensating for what they are bad at (accuracy, recency, admitting uncertainty, and resisting the gravitational pull of majority opinion).

Learn these techniques, practice them, and you will find that your research is faster, broader, and more rigorously stress-tested than before. Just remember: the LLM is the research assistant. You are still the researcher.