1
📬
Level 1 · Reading nurse
Checks your inbox
"At 8am, summarize last night's emails, and flag anything urgent." Reads, categorizes, summarizes — sends nothing on its own.
Superhuman AI · Shortwave · Gemini in Gmail
New wing · MRI & ECG department
Deep scan. Acting agent. Good prompt.
You're not here to chat. Here, the AI does research — writing a dissertation from 40 sources, organizing your emails, and if you ask, ordering food on Telegram. At the end, we'll show you how to ask for it properly: prompt recipes, with copyable examples.
Dr. Deep Research · MRI Wing
A · Agent ladder
From reading nurse to chief physician.
The word "agent" gets used for all sorts of things. Some just read, others already act on your behalf. Four levels — each building on the last.
2
📅
Level 2 · Assistant physician
Manages your calendar, replies
A guest writes: "When are you free for coffee?" — the agent checks your calendar, suggests 3 time slots, and sends the reply itself. You just nod along.
Reclaim · Motion · Cal.com AI
3
🍕
Level 3 · Attending doctor
Acts on your behalf
You write on Telegram: "Order pizza for tonight at 7pm from the usual place." The agent logs into a delivery app, fills the cart, pays, and reports back when it's done.
ChatGPT Operator · Claude Computer Use · Manus
4
🏥
Level 4 · Chief physician
Runs a multi-day project
"Launch an AI newsletter: research, write, generate an image, send a test to 3 people, revise, and publish." Works independently, reports back.
Devin · OpenHands · AutoGPT's successors
B · Deep Research lab
One question. Half an hour. Dissertation.
These tools don't answer right away. They read 30–100 sources, take notes, verify, and write a structured study — with citations. A junior analyst's day of work.
OpenAI
ChatGPT Deep Research
A 20-50 page report from 30-40 sources, with citations. A junior analyst's half-hour of work on a single question.
Open →
Google
Gemini Deep Research
The deep layer of Google Search. Reads hundreds of pages and writes a structured report — exports to Workspace.
Open →
Perplexity
Perplexity Deep Research
A fast, free version exists too. Fresh sources, visual citations. The benchmark for fact-checking.
Open →
Anthropic
Claude Research / Projects
Processes PDFs, books, and document libraries in a 200K token context. Long, coherent analyses.
Open →
xAI
Grok DeepSearch
Real-time X (Twitter) and web. The fastest option for trend analysis and market sentiment.
Open →
Google
NotebookLM
Upload your own sources (PDF, YouTube, webpages) — it summarizes, answers questions, and turns it into a podcast.
Open →
C · Prompt counter
Same recipe, different seasoning.
A bad prompt: "write an email." A good prompt specifies who it's for, what tone, how long, in what format, and if possible, shows a sample. Here are 6 prompts worth trying right now.
Recipe · anatomy of a good prompt
1.
Role
"You are an SEO expert…"
2.
Task
"Write 5 title suggestions…"
3.
Context
"The target audience is entrepreneurs 40+…"
4.
Format
"In a table, in English, max 60 characters…"
5.
Example
"Like this: …" (attach a sample too!)
Golden rule: if you couldn't do the task based on the prompt, neither can the AI. Give it as much context as you would explaining it to a new colleague — better yet: upload a sample, and ask for a similar style in response. This is the biggest quality boost you get for free.
🪞
The one-word mirror
Why it's good: Shows what the AI "sees" in you based on memory and your conversations. Brutally honest and often surprising.
Prompt · copy it
If you had to describe me in a SINGLE word based on everything you've learned about me so far, what would that word be? You may write ONLY one word — no explanation, no punctuation.
💡 Tip: Follow up with: "And why did you choose exactly that?"
🧠
Memory inventory
Why it's good: Reveals what ChatGPT has noted about you (Settings → Personalization → Memory). Many people see for the first time just how much it knows.
Prompt · copy it
List, in bullet points, EVERY single thing you have stored in memory about me. Stick to facts, don't interpret. At the end, write what you think is missing but would be useful to know about me.
🎨
Sketchbook portrait (image)
Why it's good: A classic. Works with Nano Banana / GPT-image / Midjourney — especially nice if there's memory about you or you upload a photo.
Prompt · copy it
Draw the character in the image ("Me") in a loose, free, minimalist sketchbook style. On a bright white background, freely arrange full-body drawings, close-up face portraits, tiny doodles, quick sketches, chibi/deformed variants — so the page captures the character's humor, personality, and energy. This shouldn't be a character sheet, but an illustrator's personal sketchbook page: spontaneous ideas, layered drawings, notes. Use everything you know about me. Style: hand-drawn, energetic, slightly messy, artistic, full of life.💡 Tip: Attach a reference photo of yourself — the difference is night and day.
🎬
Movie poster me
Why it's good: An earlier viral trend: the AI makes a movie poster about you based on its memory. Everyone's trying this instead of the Ghibli-style trend now.
Prompt · copy it
Design a movie poster ABOUT ME based on my life, as you know it from our conversations. Give the movie a title, subtitle, a critic quote ("★★★★★ — …"), and include a "directed by" + "starring" line at the bottom of the poster. The style should be a 90s art-house movie poster, painted, cinematic — NOT generic AI-photo-with-text-slapped-on. Text in English.🔮
Numerology & horoscope diagnosis
Why it's good: Not to be taken seriously, but incredibly entertaining — and often a surprisingly on-point reflection of your mood.
Prompt · copy it
My birth date is [YYYY.MM.DD], my name is [FULL NAME]. Produce ALL of the following in one go: (1) a numerology life-number analysis, (2) a Western horoscope profile, (3) a Chinese zodiac characterization. At the end, in a fourth paragraph, write: if all three systems agree on something, what's that common thread? Write it in a direct tone — as if a wise friend were saying it over coffee.
📁
"Projects" — your own practice
Why it's good: ChatGPT and Claude Projects: you upload your own material (PDFs, sample documents, your style), and from then on the AI ALWAYS answers the way you want. It remembers within a project.
Prompt · copy it
Start a new Project called "Newsletter Writer." Upload: 5 of your past newsletters, your brand voice description, your target audience. The instruction: "ALWAYS write in this tone. Never use clichés. Every newsletter has 3 parts: hook, story, CTA." After that, all you need to write is: "new newsletter about topic X" — and it's done.
💡 Tip: The project = your own specialized AI. Set it up once, use it many times.
Closing remarks
A good question is worth more than a good model.
A weak AI with a good prompt gives a better answer than a frontier model with a bad prompt. That's the essence of the MRI wing: what you ask, how you ask it, and who you give the task to — a nurse, an attending, or a chief physician.