Journey · Scientist · 5 full days · Level 1 to Level 3, with a look at Level 4
What you get out of the week
On Monday you arrive with one task from your research. On Friday you show a result that works, that a colleague has rerun, and that you can explain in five lines. In between you use AI where it helps and check it every time it answers.
The task stays the same all week. Each day asks a bit more of it:
- Level 1: it works for you.
- Level 2: a colleague can use it.
- Level 3: your group can rely on it without you.
Who this is for
You work in a life-science lab: PhD student, postdoc, technician or group leader. You have used a chatbot. You have never opened a terminal. If your Thursday build needs one, its book walks you through every click.
You bring one small, real task from your own work that produces numbers or a result you can check. For example:
- counting cells in 40 images
- analysing a qPCR plate
- predicting the structure of a published protein
- turning 30 protocols into one comparison table
Can't use your own data this week? Bring a published dataset from your field that looks like yours. Master's students have their own journey, which starts from a prepared lab case: Life-Science Master's Student.
The method behind every day
Monday to Thursday, each day is one trip round the four quadrants from T01-L01 · From Idea to Working Result: The Four-Quadrant Method. The quadrants form a grid of four boxes: two about the problem, two about the solution. Friday checks the result and hands it over.
| The problem | The solution | |
|---|---|---|
| Define | Structure it | Choose tools |
| Explore and test | Understand it | Build it |
- Structure it: say in one sentence what goes in, what comes out, and how you'll know it worked.
- Understand it: find out how the task is done today and what you can reuse. When that changes the problem, go back to Structure it. This back-and-forth is where most of the value is.
- Choose tools: start with the simplest option and only move up when it fails. First, something that already does the job. Then a small change to it, then a template or spreadsheet formula. Last, AI for one part only.
- Build it: make it, try it on cases where you know the answer, and fix what breaks.
How Monday to Thursday run
| Time | Quadrant | What happens |
|---|---|---|
| 09:00–10:30 | Structure it | A 20-minute demo, then you write today's problem sentence. |
| 10:45–12:15 | Understand it | You look at what already exists: your data, the papers, other people's tools. |
| 13:15–15:15 | Choose tools | You try options on known answers and pick one. Then you start building. |
| 15:30–17:00 | Build it | You finish today's result. Your partner checks it, and you write the exit check. |
Breaks: 15 minutes mid-morning, an hour for lunch and 15 minutes mid-afternoon. There's no homework. All reading happens in the room: the trainer points you to the parts of each book you need that day. The whole books are for later.
You work in pairs. Your partner is your first checker, and you check their work, too. The course runs with up to 12 people per trainer.
The five-line exit check closes every day, Friday included. It's the only paperwork:
- Problem: today's final Structure it sentence.
- What exploring changed: what you learned in Understand it, and how it changed the problem.
- Choice: the tool or method you picked, and why it's the simplest that works.
- Result: what you built, and whether it worked when your partner tried it.
- Next step: the one thing you'd improve on the next trip round.
From Monday you also keep an AI use log, one line per AI step: date, tool, what you asked, what you kept. On Friday it becomes your disclosure.
The week at a glance
| Day | Question of the day | Level | You leave with |
|---|---|---|---|
| Monday | What exactly is my problem, and what may I paste? | 1 | A problem sentence, your data sorted by colour, a prompt you can reuse |
| Tuesday | What is already known? | 1 | A checked list of sources |
| Wednesday | Which tool gets my lab data right? | 1 → 2 | A tool tested on known answers, and a first result on your data |
| Thursday | Can someone else rerun it? | 2 | A script, pipeline or small tool your partner reran with the same numbers |
| Friday | Can my group rely on it? | 3, with a look at 4 | A checked result shown live, your disclosure and one rule for your lab |
Monday: your question, safely
Before any tool, you decide what your problem really is, and which of your data may go into which AI tool.
| Quadrant | What you do | Read |
|---|---|---|
| Structure it | Write your task as one sentence. Then sort your files by colour (see below). | T12-L01 · Decide What You May Paste Into AI at Work |
| Understand it | Find out how the task is done today, and what you already have that you can reuse. Meet the four facts about AI assistants (see below). Rewrite your sentence. | T01-L01 · From Idea to Working Result: The Four-Quadrant Method |
| Choose tools | Go from the simplest option upwards. Only one part of the task goes to AI, and only if nothing simpler works. | T01-L01 · From Idea to Working Result: The Four-Quadrant Method |
| Build it | Write one prompt for that part: your material, who the answer is for, one example, what to leave out. Run it, check every claim and save the prompt. | T02-L01 · Get Answers You Can Use from a Chatbot, T03-L01 · Write a Prompt You Can Reuse |
The three colours:
- Green: already public, like published papers or protocols from a journal. Any tool.
- Yellow: internal but not sensitive, like a lab how-to with no names in it. Only tools your institute has approved.
- Red: unpublished results, anything about a person, and anything confidential. No online AI tool, unless someone who may decide has cleared this exact use in writing. This week there are no such clearances, so red stays offline.
Most lab data is red until it's published. The week's way around this comes from the books. You show the AI the shape of your data, such as the column names and a few made-up numbers, and let it build a tool. Then you run that tool on the real data on your own computer.
The four facts about AI assistants:
- It writes what sounds right.
- It starts every chat from zero.
- It sounds just as sure when it's wrong.
- What you type goes to the provider.
What goes wrong today: the assistant gives you a confident answer that turns out wrong: a reference that doesn't exist, or a number it made up.
Tools: the chat assistant your institute has approved, for example ChatGPT, Claude, Microsoft 365 Copilot or Langdock.
Tuesday: what is already known
Most lab tasks have already been solved, half-solved or warned against by someone. Today you find out which, and you check every source an AI hands you.
| Quadrant | What you do | Read |
|---|---|---|
| Structure it | Turn your problem into three questions for the literature. For each one, write what a good answer would let you do. | T17-L02 · Frame a Problem Your Team Agrees On, T15-L01 · Use AI in Your Research and Say Exactly How |
| Understand it | Search with an AI research tool, then check every hit in PubMed. Does the paper exist, and does it say what the summary claims? | T17-L03 · Explore What Already Exists Before You Build |
| Choose tools | Decide where your papers go. Published papers are green, so any notebook tool will do: a tool that answers only from the papers you upload. Your own notes follow their colour. | T04-L01 · Ask Your Own Documents and Check Every Citation |
| Build it | Load 5–10 papers and ask your three questions. Mark each claim in the answers as supported, unsupported or not found. | T04-L01 · Ask Your Own Documents and Check Every Citation; optional: T04-L02 · Curate a Document Set Your Team Can Ask |
What goes wrong today: a summary cites a real paper, but the paper says the opposite. You only catch it because you opened the paper.
Tools: Perplexity to map the field, PubMed to check it, and Gemini notebooks for your papers. DeepL handles papers in other languages.
Example: Chat over documents.
Wednesday: your lab data
Today a tool works on your own data. The rule of the day: before you trust a tool, test it on cases where you already know the answer. Pick one of three tracks.
| Quadrant | What you do | Read |
|---|---|---|
| Structure it | Write down at least ten known-answer cases: images you counted by hand, a plate you already analysed, a structure that is already solved. | T16-L03 · Test a Tool on Your Own Work Before Your Team Depends on It |
| Understand it | Look at your data the way the tool will see it: file types, empty wells, clumped cells, blurry images. Add the awkward cases to your list. | Your track's book |
| Choose tools | Try two options on the same cases, and keep the one that gets them right with the least effort. | T16-L01 · Pick the Right Tool for One Task by Trying It |
| Build it | Run the chosen tool on your real data, and compare ten results with your hand values. | Your track's book |
The three tracks:
- Images: count cells with a ready-made model that runs on your computer, and check the counts by hand.
- Read T18-L01 · Count Cells With a Ready-Made Model and Check It by Hand.
- Tools: Cellpose, with counts checked in Fiji.
- Example: Segment cells with Cellpose.
- Structures: predict a protein structure and read its confidence scores honestly.
- Read T18-L02 · Predict a Protein Structure and Read Its Confidence Honestly.
- Tool: AlphaFold Server. This week, published sequences only: the sequence goes to Google, and an unpublished one is red.
- Example: Explore molecular structures.
- Tables: ask a spreadsheet a question and check which rows the answer used, then clean a messy instrument export once.
- Read T10-L01 · Ask a Spreadsheet a Question and Check Which Rows It Used.
- Then read T10-L02 · Clean a Messy Export Once, So Your Team Can Refresh It.
- An unpublished plate is red. Ask the assistant with made-up numbers in the same layout, then run the formula it gives you on the real file in your own spreadsheet.
What goes wrong today: the tool gets your easy cases right and fails on the awkward ones you almost didn't test, such as clumped cells, a failed well or a floppy loop in the protein. This is why the test cases come first.
Thursday: make it rerunnable
A result only you can produce is a result nobody can check. Today you turn Wednesday's work into something your partner can run on their own laptop, getting the same numbers.
| Quadrant | What you do | Read |
|---|---|---|
| Structure it | Write what your partner needs to rerun it: which input files, which tool version, which prompt and which settings. | T15-L02 · Record an AI Step So a Co-Author Can Rerun It |
| Understand it | Check the colour of every file your build touches. For a red file, you have two options. One: the AI only ever sees made-up numbers. Two: the AI step runs on a model on your own laptop. | T11-L01 · Run an AI Model on Your Own Laptop, Offline |
| Choose tools | Pick the build that fits your track (see below). A coding assistant builds it with you, one small change at a time. | Your build book |
| Build it | Build it, then hand it to your partner. They run it from your notes alone and compare the numbers. | Your build book |
Pick one build:
- A small tool that runs in your browser, for example a Ct calculator, or a page that shows the prediction's confidence along your protein: T06-L01 · Build a Small Browser Tool You Will Actually Use.
- A script that repeats your analysis on each new export: T08-L02 · Get a Script Your Colleague Can Rerun, Without Being a Developer.
- A counting pipeline for a whole folder of images: T18-L03 · Build a Cell-Counting Pipeline Your Group Can Rerun and Check.
- An analysis that gives the same numbers every time it runs: T10-L03 · Build an Analysis Your Colleague Can Rerun and Get the Same Numbers.
Same numbers, even with AI in it. An AI can answer differently each time you ask. So either the AI only writes the tool and the finished tool contains no AI, or you save the AI's output with your notes and your partner checks that.
If your data is about people (participants, patients, interviews), it's red, and it needs a formal review before any AI step. Bring a published dataset this week. Then read T15-L03 · Get an AI Step on Participant Data Reviewed, Then Build It Inside the Approval before you bring your own data in.
Your partner reruns it on a small sample you're allowed to share, or on made-up data with the same layout.
What goes wrong today: your partner can't get your numbers, because a tool version, an input file or the exact prompt was missing from your notes. You fix the notes, not the partner.
Tools: opencode as the coding assistant (the Desktop app, where you describe what you want and it writes the code), and LM Studio for a model that runs offline.
Friday: hand it to your group
Today you find out whether your group could rely on your result without you in the room. You check, fix, show and look ahead, and your fifth exit check covers the whole week.
| Time | What happens | Read |
|---|---|---|
| 09:00–10:30 | Swap results with another pair. Check theirs the way a careful colleague would, and write them a short check note. | T01-L02 · Check AI Output Before Your Team Relies on It |
| 10:45–12:15 | Fix what the check found. Remove whatever nobody needs. Turn your use log into a disclosure: which tool, which step, what you checked. | T01-L03 · Build the Smallest Thing Your Colleagues Can Run Without You, T15-L01 · Use AI in Your Research and Say Exactly How |
| 13:15–15:15 | Show your result, 8 minutes each: run it live, read your five lines, answer two questions. | — |
| 15:30–16:15 | A look at Level 4: what it would take to set up AI for your whole research group. | T15-L04 · Set Up AI for a Whole Research Group |
| 16:15–17:00 | Write one rule for your lab, on one page, that a new colleague could follow without asking you. Then write your fifth exit check, on the whole week. | T13-L03 · Write a One-Page AI Rule Your Team Can Use Without Asking You |
During the Level 4 look you also see Heidelberg AI Desk. It runs Open WebUI, a chat window, on one computer, so a group can try a shared chat with a local model before anyone asks for a server.
How you pass
There are no points. You pass when all five of these are true:
- It runs. You show your result live on Friday.
- Someone else reran it. Your partner got the same numbers from your notes alone.
- It was checked against known answers: at least ten, including the awkward ones.
- Your paperwork is complete: five exit checks, the use log and your disclosure.
- No red data went into an online AI tool. Not once, not "just to try".
What you take home: your result, your notes for rerunning it, five exit checks, a check note from another pair, your disclosure and one rule for your lab.
Before Monday
- Pick your task and get your group leader's OK to use its data. If in doubt, bring a published dataset that looks like yours.
- An account for your institute's approved chat assistant. If there isn't one, the course provides one, for green material only.
- A laptop you may install software on. For an institute laptop, ask IT a week ahead. It should be a Mac with an Apple chip or a recent Windows PC, with 16 GB of memory for the offline model (8 GB works with smaller models) and about 10 GB of free disk space.
- The install list. Your trainer sends it a week ahead: LM Studio, opencode Desktop, and Fiji for the images track. It takes about 30 minutes.
- Nothing to read in advance.
What this week leaves out
- Setting up servers.
- Fine-tuning models.
- Automations that run on their own.
- AI on clinical or patient data.
Each has its own book at a higher level, for when you need it.
Further reading and tools
All the books in this journey are listed below the course, in the order you meet them.
- Used hands-on: Cellpose, Fiji, AlphaFold Server, PubMed, Perplexity, Gemini, LM Studio, opencode.
- Mentioned: DeepL, ChatGPT, Claude, Microsoft 365 Copilot, Langdock, Heidelberg AI Desk, Open WebUI.
Rule to remember: one task, round the quadrants again and again, and nothing counts until someone else has rerun it.