A Writer's Workflow To AI Research Without Losing Your Ability To Think

Every honest writer will tell you about the research that they put into their work. Many writers have a vested interest in ensuring that their work is grounded in reality because they are incentivized to do so. From the biggest of breaking news stories to the smallest of fiction pieces, mankind wants the written works they consume to be mapped to reality because that is how we engage with the world: as stories.

Before the Internet, this research was manual. As a writer, you would take your experiences with other people, the information you found in papers and interviews, and all manner of resources and combine them into your work, whatever it may be. Newspapers have their sources, novelists have their folklore and mythology, and so on.

For the last few decades, the Internet has made it possible to obtain information at a speed and scale not previously seen. Tens of thousands of words can be accessed so long as you have a connection, a capable device, and a will to learn.

I mention all this because it's no secret that LLMs and other AI tools are here to stay. I believe these tools, like search engines before them, will continue to integrate into software and workflows as the years go on. Though I don’t believe that they will be the downfall of mankind, I do believe there will be growing pains along the way, and it is worthwhile to get ahead of those disruptions to maintain the craft of writing in a way that is holistic between man and modern tech.

This approach starts as the genesis of writing: the research step mentioned earlier. In a world where information can be sourced, summarized, and deposited for you in seconds by a fallible piece of software, there are rules and frameworks writers need to maintain to stay ahead of the second-order issues caused by LLMs.

Nobody became a developer to write documentation. But the docs still need to get written — PRDs, README updates, architecture decisions, onboarding guides.

Wispr Flow lets you talk through it instead. Speak naturally about what the code does, how it works, and why you built it that way. Flow formats everything into clean, professional text you can paste into Notion, Confluence, or GitHub.

Used by engineering teams at OpenAI, Vercel, and Clay. 89% of messages sent with zero edits. Works system-wide on Mac, Windows, and iPhone.

Try Wispr Flow free

Before delving into the steps I recommend for writing research, I think it’s important to explain what to avoid during this practice upfront. I don’t want anyone using LLMs to help them research information to have the framework without knowing the restrictions, because there are dangers when using these burgeoning tools.

This is the first rule because it is the most important specifically for using AI tools of any kind. While these tools raise the skill floor on many different kinds of tasks, it doesn’t change the fact that these tools guess when trying to assist you based on what the aggregate or standardized answer would be for your task. Without your unique context or understanding of the matter, these tools will offer you something generic and half-baked, all while claiming their answer is the most correct.

For knowledge work like writing, this is a death knell for your output. Not because you won’t have words to present, but because your output will look amateurish, unresolved, and/or internally inconsistent.

Don’t let that happen. If you can’t create a longform piece on your own without guidance or assistance from an AI tool, you haven’t done that task well enough to pass it off yet.

These tools are like the overly confident, underskilled office intern hopped up on too many TikTok manifesting videos. They don’t know how full of themselves they are until you correct them, and until you do, they’ll believe they are heaven’s gift to you.

You can’t correct them without your own expertise to push back, so develop that first before going to these tools.

I hope it goes without saying that, when doing research, you want access to the sources of your information, regardless of how you obtain them. Research papers, articles, videos… all of these sources should be available to you so that you have the ability to go in and review these sources for yourself in spite of AI tools being able to provide high-level summaries for them.

Firstly, LLMs will pull from unofficial sources if you don’t explicitly prevent them from doing so in your prompting or in your settings. A biomedical student learning about the physiological pathways of vaccines could benefit from clinical research papers, but not the ramblings of a prolific almond mom on Facebook. AI provides answers based on popular, but not always accurate, information without direction.

Even if you do prevent any unwanted sources from entering into the equation, there is still a chance you won’t care for the sources LLMs can pull for you. The tool might ignore a methodology flaw, or grab research backed by a biased benefactor that the tool doesn’t know is biased, but you do.

I’ve written before about why curating your own knowledge vault is important, and this reality of AI research patterns further proves those points. You need to be able to not just get information, but parse through it for yourself and piece it together with other sources you find. Those interconnections are where some of the magic in knowledge occurs, and it’s a skill you can’t hone if the tool does it for you.

Regardless of what the tooling dredges up for you, whether it’s worth it or not, you should not outsource your communication of those results to other humans via your LLMs or agents.

There is research showing that, in their current incarnations, LLMs communicate differently than humans. Specifically, the ways in which humans and LLMs share stories differ greatly.

The above is an analysis of thousands of stories written by humans compared to those written by different AI tools, provided by Russell et al. in their 2026 paper “StoryScope: Investigating idiosyncrasies in AI fiction.” In this paper, the researchers found that humans introduce things that AI tools don’t, like emotional context, ambiguity, moral greyness, and other nuances that these tools cannot currently replicate.

It makes sense if you think about it. These models are designed to give repeatable, consistent answers, sourced by the vast range of materials they’re trained on. You can’t be consistent while also breaking out into tangents randomly.

So, this research partly explains why human-to-human conversation feels better than human-to-AI: there’s a messy, emotional layer and flexibility in the temporal structure of human conversation that we’re better attuned for. That’s why getting an email from your agent or one drafted by ChatGPT feels worse for the reader than one sent by a human.

With the ground rules covered, we can delve into the best practices for researching using these AI tools. I have found that having three different layers of research, divided into tiers based on the level of depth and the level of visualization you need for your task, comes into play.

For a new task, I find starting at Layer 1 and working one’s way up through the layers makes the most sense. Of course, if you already have a baseline that covers one or more of the lower layers, by all means skip ahead to the layer most appropriate to you.

The first layer involves creating a space where you can create many threads based on ideas you want to explore or get feedback on from these tools. I like to call this category “Brainstorm Buddy” due to its focus on asking questions to help you clarify your thoughts, and because I’m a sucker for alliteration.

I’m including prompts for each of these layers, if you decide to make something like this for yourself. For the sake of brevity, though, all of the prompts will be hyperlinked, so if I reference specific lines or concepts in the prompt, you’ll find that context in the hyperlink itself.

Anyways, you’ll grab the Brainstorm Buddy prompt and set up your project, or custom GPT, or whatever your tool’s version of categorizing prompts is. Plug in this prompt as the project instructions, and then make a new thread and start dumping your thoughts and ideas into the project.

In each thread, the Buddy will go through and ask clarifying questions about how you expect the idea or thought to work, and how you arrived at that conclusion.

Say, for example, you wanted to start a Substack on your field of expertise. You would go into the Brainstorm Buddy and tell it what you want the Substack to be about, what your specific views on the topic are, and how you expect the content to work. It will ask questions, you’ll answer those questions, and keep going through several rounds of questions and answers.

That will repeat until reaching a point where everything is laid out for you, based on your thoughts and some nudges from the LLM. For the most part, this takes four to five passes before you hit the end of the thread. If it takes longer than that, that can be okay, but ideally you don’t want to go past eight to ten messages that you send off to the Brainstorm Buddy per idea.

The idea here is that creatives can get stuck in their minds about their ideas, and rather than let those ideas get stress-tested against another source of opinion, we’ll sit on those thoughts and do nothing with them, like self-flagellating ourselves with wandering minds is a noble act.

It isn’t. An idea that can’t be backed up by guiding principles and clear outcomes won’t take you far. This layer helps you get past that initial hurdle.

The Research Assistant layer is a prompt meant to take a well-reviewed idea and turn it towards the existing data available so that viability and best practices make themselves known from the outset.

Much like before, you’ll grab the Three Player AI Research Workflow below (don’t worry, it’s free) and make a new project using your tool’s version of that setup.

Once set, you can take your findings from the Brainstorm Buddy and bring it over into this new project, where you can ask about market fit, competitors and comparisons, and all other sorts of quantitative metrics that let you see how your idea stacks up to what’s already out there.

Specifically what you want from your brainstorming thread is the

I’d be sure that whatever sort of “extended thinking” or higher-end model options you have access to are enabled for this thread, at least for the initial buildout of the thread. Additionally, while it’s already part of the prompt structure, I would turn off any sort of social media sourcing your LLM would otherwise do to be sure that you are more likely to get hard data, and not someone’s random account of it in your outputs.

From there, you have three follow-ups you’re looking to ask the initial data:

Going through this list of follow-ups means that you now understand not just what the data is telling you about your idea, but also what risks exist in its implementation, and what work you need to do next to overcome or mitigate those risks.

Once you have those follow-up questions answered, you’ll want to grab the entire chatlog from both the Brainstorm Buddy and the Research Assistant as you go into Layer 3.

The next layer covers the drafting of plans, artifacts, or other assets you can use to get your idea off the ground. If you don’t have the ability to create files or to visualize data using your tool of choice, this section won’t be helpful to you directly, but it does offer a look at what’s possible if you need more than just text on a screen to get started.

Same process to start this up as before: grab the prompt instructions, toss it into a new project, and start plugging away. Make sure the logs you grabbed from Layers 1 and 2 go into the context for this project alongside your answers to the fill-in-the-blanks in the prompt skeleton I provided, as that will reinforce the do’s and don’ts you define for this run.

The main goal with this layer of research won’t so much be to garner new information or insights. Rather, this step is here to give you another way to piece together all the data gathered from Layer 2 and put it together for you here.

For writing, the temptation will be there to have the Modeler write a draft for you if you’re working on something.

Don’t let it. I’ve made my stance on whether AI should be writing for you or not clear, so I won’t rehash it again, but I will just remind you that the prompts are here to help you move forward on a task, not do the work for you.

Anyways, the reason I leave this part of last is that the AI tools that offer the ability to create files for you, also called artifacts, need good data to get the most out of the feature. It’s also more expensive to make artifacts than it is to use a chat window, and since I don’t expect AI credits to get cheaper anytime soon, I’d rather you be efficient with your usage from the get-go.

The above process will map onto the workload of any nonfiction writer, but the process is less obvious for fiction writing. After all, fiction doesn’t need to rely on real-world data for its beats to land.

But understanding how the real world operates can help make for an engaging story for the reader, which is ultimately our goal.

The issue presented with fiction work comes down to the moving parts one must keep in mind while writing the draft. Characters, settings, historical events for the world, and so on all need to interconnect in a way that suits the story without bloating the narrative and drowning the reader in pointless proper nouns and tedium.

So, while the above frameworks apply well to nonfiction, I’d be relying mostly on the Brainstorm Buddy to sort through ideas with fiction. Providing drafts or scenes you’ve written as part of your context can help guide your LLM tool of choice in not just the context of your story, but also how you like to write.

The Research Assistant and Junior Modeler prompts will be handy for any research you need to do on topics within your writing, or inspiration that you want to aggregate into one place.

Much like in the practice of writing fiction itself, a novelist would need to get creative with these prompts to get the most out of them.

Many people fear that utilization of AI tools will make people less intelligent. It’s a valid fear, especially since the data suggests that’s what happens right now with the average user.

The problem is that this is a function of how the tool is provided to anyone with an email and a few minutes to sign up for the software. It also doesn’t help that there are people out there talking about how to get results with AI, but not how to get results that don’t hurt your cognitive function.

I believe the biggest takeaway from this entire piece, if there is one I can suggest you keep, is this: always have the tool ask you follow-up questions. Not the kinds of questions like “can I build this out for you now?”

Rather, it should be asking, “What’s the next layer above or below this work?” In other words, it should always be pushing you deeper into your own mind, rather than giving you the answer outright. That’s the entire goal of the Three Layer Research framework today.

Knowledge work will no doubt continue to change as these tools proliferate further. My hope is that, with these rules and frameworks offered above, you can better ready yourself to keep your mind sharp in a world that continually offers excuses to dull yourself instead.

Dictate acceptance criteria and reproductions inside Cursor or Warp. Wispr Flow auto-tags file names, preserves syntax, and gives you paste-ready text in seconds. 4x faster than typing.

Try Wispr Flow free