Give Your AI an Upgrade: Four Levels Every Researcher Should Know
Updated: 5 days ago
After teaching AI to researchers for two years, it’s clear that most people are stuck using AI as a glorified Google, and that’s a real shame because it can do so much more!
I've been thinking about this a lot while putting together the new Advanced AI Usage module for AI for Researchers, and it's settled into what I'm calling the AI Implementation Ladder. It's four levels, and almost everyone lives entirely on level one. That's not a criticism. It's where most people start, and you just don't know what you don't know. But there's a lot more value sitting above it, and the good news is that none of it requires you to become a developer.
Here's the ladder, bottom to top, with real examples from my own work at each level.
Level 1: AI for answers
This is AI as a glorified version of Google. You ask a question, you get an answer. Every consumer tool does this out of the box: ChatGPT, Claude, Gemini, Copilot, all of them, no setup required.
It's genuinely useful for quick jobs like generating a title, making a paragraph longer or shorter, tightening up formal language, or sketching a travel itinerary. Most researchers I talk to use AI almost exclusively at this level, and I get why. It works, and requires zero setup.
The problem is that the answers are generic. You get the same response anyone else asking the same question would get, because the AI doesn't actually know you. Having chat history or a "memory" feature turned on doesn't count as intentional context. That's passive. Level 2 is the step change you want.
Level 2: An AI that knows you
This is where you deliberately engineer context so the AI behaves differently for you than it would for anyone else, including writing in your voice.
Set up your unique system prompt
Every AI company applies a system prompt on top of the raw LLM, which tells the AI how to behave and you cannot touch it. But few people realise that there is a second “system prompt”, which you can change and personalise at will. The easiest way I've found to build your own system prompt is Socratic prompting. In short: ask the AI to interview you about who you are, your research, how you like to work, what you're trying to achieve and more. I recommend you use dictation so it feels like an actual interview. Then get it to summarise that into a permanent instructions document.
My own Notion AI (I call her Clippy) has a very detailed profile covering my role, working style, the people I work with, my background, my goals and more. Gosh, it even knows that I am into salsa dancing and scuba diving! Is that too much? Importantly, I also gave it instructions on my writing style and tone, in a way, my voice. I built that voice doc by feeding it old blog posts that performed well into the AI and asking it to analyse the tone, sentence structure, and word choices, then turning that into a short style guide it now checks against. Does it work? Well, I would say it's not my clone, but it's certainly better than the default factory settings.
Once you've got that doc, paste it into the customisation settings of whichever AI you use. Here's where to find them:
ChatGPT: Settings → Personalisation → Custom instructions
Claude: your initials (bottom left) → Settings → Instructions for Claude
Gemini: Settings → Saved info → Your instructions for Gemini
Copilot: Chat settings → Personalisation → Custom instructions
These menus get reshuffled every few months, so if yours looks different, search the settings for "instructions" and you'll find it.
Connect your AI with other platforms you use
AI is a marvellous invention, but what's not so sexy is copy-pasting or exporting/importing CSV files across platforms all the time. Here's where connectors come into play, which plug your AI into other tools, usually through an API or the newer MCP standard. Connectors let the AI reach into your email, calendar, Drive, Zotero, or a Notion/Obsidian workspace, so it's accessing your actual material instead of guessing. This, for me, has been a big unlock because with the right connector, you dramatically expand the usefulness and range of applications of your AI.
Let me give you a couple of examples. Connecting Claude to Google Analytics has been incredible for understanding how people behave on our website in a way I would have never been able to figure out on my own. It's literally like having an expert data analyst on tap!
For researchers, the number one use case I can think of for a connector is linking your general-purpose AI, like ChatGPT or Claude, with a scholarly literature database. Consensus and Elicit are good ones to start with as they have an MCP.
The limitation at this level is that you're still the bottleneck. Nothing happens until you sit down and prompt it, and that's why you want to venture to level 3.
Level 3: An AI that works without you
Set up an agent on a trigger or a schedule and you're no longer the one kicking things off. I've got several of these running for AYS right now, but here is my favourite.
Let me introduce you to my meeting prep agent in Notion.
Here is how it works. When a new event lands on my calendar, it searches my CRM (jargon for a big spreadsheet of my contacts) and inboxes, runs a web and LinkedIn search on the person I'm meeting, and drops a summary into the CRM before I walk into the meeting.
As somebody with a pretty crappy memory, this is extremely useful because I often forget about previous interactions I had with a person. I found myself in plenty of awkward situations reintroducing myself to people I should have remembered 😬.
A version you can use: the weekly literature agent
That meeting prep one is fairly business-specific, so here's the level 3 agent I'd build first if I were still doing research.
Every Monday morning it searches PubMed and arXiv for everything published in the last seven days in your field, compares each result against a short written description of your actual research question, and emails you the three papers worth reading, with one line on why each one matters to you. Everything else gets binned.
You might be thinking that PubMed and Google Scholar alerts already do this. Not quite. Those match keywords and dump the lot in your inbox, which is exactly why most of us stop opening them after a month. The difference is the filtering step: the agent reads the abstracts and judges relevance against your research question, so what arrives is a shortlist rather than a pile. You can also tell it what you are not interested in, which is half the battle.
Setting it up is a scheduled prompt, not code. You write your research question and inclusion criteria once, point the agent at the databases, set it to run weekly, and tell it where to send the output.
One caveat: it will occasionally bin something you would have wanted. Treat it as triage that keeps you current week to week, not as a replacement for a proper systematic search.
The limitation of level 3 is that it works only for standardised, repeatable tasks. It's no good for anything that needs judgement on the fly, but for the repeatable stuff, work gets done while I'm off doing something else entirely.
Level 4: AI coding
Get ready to venture into mind-blowing territory! Here we are using tools like Claude Code, ChatGPT Codex, or Lovable to build actual software, without being a developer. It's the level I was most sceptical of, and it's now the one that I love the most.
In less than a week of actual work, I built three things without writing a single line of code.
The first is The Augmented Researcher platform, a directory of AI tools for research with reviews, pricing, and data policy on each card, a news section fed automatically by an agent, and a collection of AI training programs for researchers. Notion is the backend, and it syncs across on its own. It will be live before the end of the year.
The second is a Typeform replacement for internal use. It is an on-brand form tool with better analytics, including Net Promoter Score reporting, because Typeform kept getting more expensive and I kept hitting its limits.
The third is an interactive dashboard for my old PhD research. It turned my ocean acidification manuscript, underwater soundscape recordings, and the accompanying dataset into something you can actually explore: audio playback, a 3D plot of the soundscape over time, a map of the field sites, sliders that update the graphs live, and a plain-English "so what" section on what it means for fish stocks. I think this is a very interesting format that provides a new avenue for research dissemination beyond the classics of graphical abstracts and video abstracts.
This last one took under two hours from start to finish, with a first working version after about 40 minutes. As an agency owner I can tell you that a developer quote for the same thing would likely run past $10,000 and take weeks.
But before you get too excited and go off building God knows what, there is a caveat here that matters: this is fine for low-risk projects like personal sites or tools that don't collect user data. For anything touching payments or personal data, I still want an actual developer to review it before it goes anywhere near real users.
Where you probably sit right now
If you're only ever asking ChatGPT quick questions, you're at level 1, and that's a perfectly good place to start. I started there too! The jump from level 1 to level 2 takes a little bit of setup, and that's exactly why most people don't bother. But the payoff is immense. Once the AI actually knows you and has appropriate context, larger and more complex tasks stop feeling like magic tricks and become feasible.
I walk through all four levels in a lot more depth, including how to build your own context document and set up your first agent, in the new Advanced AI Usage module. It's included in both the Essential and Premium tiers of AI for Researchers Cohort #2, and if you've already purchased the course, it is already in your account.
So, where do you reckon you sit on the ladder right now?
I'd genuinely like to know. Ping me on LinkedIn and tell me which level you're on, and which one you're going after next.
