AI in the terminal
The handout · Department of Finance · 16 October 2026
How to use this
Everything from today is in here, including the parts we don’t get to in the room. Every grey box with a coral edge is a prompt: click the copy icon in its corner, paste it into Claude Code or Codex, and change the bits in [square brackets].
The shape of the day: the morning gets you running, has the agent do something tedious for you, and then takes your paper apart. The afternoon blocks (sections 5–8) work in any order, today or next week.
| Section | What you get | |
|---|---|---|
| Morning | 1. Get it running | Claude Code or Codex, and git, working on your laptop |
| 2. Do something tedious | A messy folder cleaned up, with a save point and an undo | |
| 3. What it knows about you | A profile that makes every answer fit you | |
| 4. Take a paper apart | Its biggest problem, and an attempt at a fix | |
| Afternoon, pick | 5. Referee panel | Simulated referee reports, and an editor’s letter |
| 6. References | Searched, downloaded, DOI-checked, claim-checked | |
| 7. Leave a trail | A folder a co-author can rerun | |
| 8. Notes and structure | Where things go; a project file; notes the agent keeps | |
| Any time | 9. Your own work | Starting points for next week |
| 10. GitHub | A copy of the folder in the cloud; co-authors | |
| 11. Connectors | Your email, calendar and files, if you want them |
Two rules for the day. Don’t overthink it: just use it. And when you get stuck, ask the tool. “How do I…?”, “What just happened?”, “Undo that.” It is its own best manual.
1 Get it running
1.1 What you need
A paid subscription. You only need one:
| If you pay for… | …install |
|---|---|
| Claude (Pro or Max) | Claude Code |
| ChatGPT (Plus, Pro, Business, Edu, Enterprise) | Codex |
Free plans don’t work for either tool. BI’s Gemini subscription doesn’t either: it can’t touch the files on your computer. No subscription today? Pair up with your neighbour, and sort one out tonight. One month of Claude Pro or ChatGPT Plus is plenty to start.
No admin rights needed for the tools themselves; both install into your own user folder. Git (step 5) may ask for admin rights, and there’s a way around that.
1.2 Install on Windows
Step 1: open PowerShell. Click Start, type powershell, and click Windows PowerShell. Not the one ending in (x86). You’ll see a line like PS C:\Users\YourName>. To paste into PowerShell, right-click in the window or press Ctrl+V.
Step 2: install your tool. Paste one of these and press Enter.
Claude subscription (Claude Code):
irm https://claude.ai/install.ps1 | iex
[Environment]::SetEnvironmentVariable("Path", [Environment]::GetEnvironmentVariable("Path","User") + ";$env:USERPROFILE\.local\bin", "User")Both lines, in one paste. The second puts claude on your PATH, so Windows can find it. (No admin rights needed: it only changes your own settings.)
ChatGPT subscription (Codex):
powershell -ExecutionPolicy ByPass -c "irm https://chatgpt.com/codex/install.ps1 | iex"Wait 1–3 minutes until the PS C:\...> prompt comes back. Then close PowerShell completely and open it again. That step matters: it lets Windows find the new program. Check:
claude --version(or codex --version). A version number means you’re installed.
Step 3: log in. Type claude (or codex) and press Enter. Your browser opens: log in with the account your subscription is on, approve, and go back to PowerShell. Codex: choose Sign in with ChatGPT. You only do this once.
Step 4: the test. With the tool running, type a sentence:
What files are in this folder?If it answers in plain English, you’re done. To quit, type /exit (or press Ctrl+C twice).
Step 5: Git (everyone). Git lets the agent save a snapshot before it changes anything, so you can always go back. You’ll use it from the first exercise on. You never type git commands; you ask the agent. Quit the tool first (/exit), then:
winget install --id Git.Git -e --source wingetClick Yes when Windows asks. Close PowerShell, open it again, and check with git --version.
Can’t click Yes (no admin rights)? Start the agent and ask:
Install the portable version of Git for Windows in my user folder, without admin rights, and put it on my PATH. Then check that git works in a new PowerShell window.wingetnot recognized? Download the installer from https://git-scm.com/downloads/win and click Next on every screen.
1.3 Install on a Mac
Open Terminal (Cmd+Space, type terminal, Enter) and run one of these:
# Claude Code
curl -fsSL https://claude.ai/install.sh | bash
# Codex (needs Homebrew, from brew.sh)
brew install --cask codexThen log in and test as in steps 3 and 4. For git, type git --version. If your Mac offers to install the command line developer tools, click Install.
1.4 If something goes wrong
Work down the list. The first two fix almost everything.
claudeorcodexis not recognized. You skipped “close and reopen PowerShell”. Open a fresh window and try again. Still failing? Run the install line again and watch for red text.Claude Code: let it diagnose itself. Run
claude doctor. It often names the problem outright.Codex stops and asks for administrator approval, or about a “sandbox”. Choose to continue without elevated setup (the restricted or “unelevated” mode). Codex works fine that way. This is what happened on BI laptops in August.
irmis not recognized. You’re in Command Prompt, not PowerShell. Open Windows PowerShell as in step 1.The download fails, hangs, or says it’s blocked. That’s BI’s network or security software, not you. Put your hand up. For IT, the request is: allow
claude.aianddownloads.claude.ai(Claude Code) orchatgpt.comandreleases.openai.com(Codex), and allowlistclaude.exe/codex.exein my user folder.Claude Code script blocked? This often works instead:
winget install Anthropic.ClaudeCode“You need a paid plan”, or login fails. Check that you log in with the same account the subscription is on.
1.5 The terminal on one page
Five commands and a few keys, all for moving around. Everything else, you ask the agent to do.
| Type this | And it… |
|---|---|
cd foldername |
moves into a folder. cd .. steps back out. |
ls |
lists what’s in this folder (dir also works on Windows) |
mkdir name |
makes a new folder |
pwd |
prints where you are |
explorer . |
opens this folder as an ordinary window (Mac: open .) |
claude or codex |
starts the agent, right here |
| Press | And it… |
|---|---|
| Tab | finishes a name: type the first letters of a folder, press Tab. It handles spaces in names. |
| ↑ | brings back your last command |
| Ctrl+V | pastes (right-click works too; Mac: Cmd+V) |
| Ctrl+C | stops whatever is running. Twice, or /exit, leaves the agent. |
| Ctrl+L | clears the screen |
| Esc | stops the agent mid-answer, so you can redirect it |
Lost? Close the window and open a new one; nothing breaks.
Skip cd: open a terminal already in the right folder.
- Windows 11: right-click the folder (or empty space inside it) → Open in Terminal.
- Windows 10: open the folder, click the address bar, type
powershell, press Enter. - Mac: drag the folder onto the Terminal icon in the Dock. (Right-click → New Terminal at Folder works too, once you switch it on in System Settings → Keyboard → Keyboard Shortcuts → Services.)
- Anywhere: type
cd(with a space) and drag the folder into the terminal window. It types the path for you.
This is the everyday move: open the folder, open a terminal in it, type claude.
1.6 First contact
Make a practice folder, look at it, and start the agent in it:
mkdir practice
cd practice
explorer .
claudeexplorer . opens the folder you’re in as an ordinary window (Mac: open .). It’s a normal folder: anything you drag into it, the agent can see.
(or codex). Then:
What files can you see here? Do not change anything.The floor: ask it to make a file, read it back to you, and delete it again. Done? You’ve used an AI agent on your own machine.
If you’re early, find the edge. Pick whatever tempts you:
Something from finance.
Price a European call with S = 100, K = 105, r = 4%, sigma = 20%, T = 1. Show the code and run it.Watch it build something. “Write a small program that does X, and run it.” It doesn’t just describe; it does.
Try to catch it out. Something very recent, an obscure figure, a specific citation. Then check the answer.
Be vague on purpose. “Make me something useful for my course.” Does it ask you questions, or charge ahead?
Write down one answer that surprised you, and one you didn’t trust, and why.
1.7 Which model, and your limits
| Claude Code | Codex | |
|---|---|---|
| Good | Opus, Fable | Sol, Astra |
| Fast | Sonnet, Haiku | Luna |
- Use Opus (Claude Code) or Sol (Codex) by default. Fable and Astra are stronger still, and very expensive: they eat your limit fast.
- Fast models are for massive amounts of repetitive, easy work. Luna is very cheap. Haiku is cheap, but not at Luna’s level.
- Switch with
/modelin either tool.
Usage limits. Every answer costs computing power; the strong models and long thinking cost the most. Claude Pro runs out sooner within a five-hour window; ChatGPT Plus runs out sooner over a week. Higher plans give many times more.
- What eats it: long sessions, big files, the strongest model, long thinking.
- See where you stand:
/usagein Claude Code,/statusin Codex. - Hit the limit? Wait for the reset, switch to a Fast model, or pair up. If you keep hitting it, a higher plan is the fix.
1.8 Talk instead of type
Speech-to-text makes long prompts painless, and long prompts get better answers.
- Windows: press Windows key + H in any text field, including PowerShell.
- Mac: press Fn twice (or the microphone key). Turn it on under System Settings → Keyboard → Dictation if nothing happens.
1.9 Working safely
- It works where you start it. Start it in a project folder, not your whole Documents folder.
- It asks before it does anything big. Read before you say yes. Ask for a plan first: “Tell me what you’d do. Don’t change anything yet.”
- Work on a copy of anything precious, and use git (section 2) for save points.
- What it reads is sent to the AI company. Treat it like uploading to a cloud service. Keep out personal data (students, interview subjects) and confidential material.
- Tired of approving every step? Codex: run
codex --full-auto, which stops asking but stays inside its sandbox. Claude Code: Shift+Tab to “accept edits”, which approves edits in the folder but still asks before anything else.
2 Do something tedious
A messy folder, on your own machine: the job a chat window can’t do, because it can’t touch your files.
The folder is an exam export, the way it comes out of WISEflow: 15 group submissions in folders named like [3-2048892] - Group 1, each with a cover sheet and an “Appendix material” folder full of files called EKSAMEN.pdf, final.pbix or dashboard_v3_FINAL.pbix, plus an empty grade sheet in a different order from everything else. It’s synthetic, so nothing real is at risk.
The floor: the folder is reorganised and committed, and you’ve undone one change with git.
2.1 Get it, and save a point
In your practice folder:
Download https://github.com/JonasMoss/vibing/raw/main/courses/2026-10-finance/props/decrap-me.zip into this folder and unzip it.Then, before it changes anything:
Set up git here and commit everything.The first time, git needs your name and email:
Set up git with my name, [Your Name], and my email, [you@bi.no].Now it can’t do anything you can’t undo. Note the order: a save point taken after the mess is a save point of the mess.
2.2 Clean it up
What's in decrap-me? What would annoy a teacher about it? Don't change anything.Reorganise it so I can grade quickly: one folder per group, files named by group and what they are, and the grade sheet sorted by group. Show me the plan first.Argue with the plan. Change at least one thing: the folder names, what happens to the cover sheets, whatever you’d do differently. Then “Go ahead.”
Make a grading table: group, participant ID, report file, dashboard file, and an empty column for the mark.2.3 What changed, and the undo
What changed since the last commit?Every move and rename, listed. Then “Commit.” Now be bold:
Delete the Marking guidance folder.Gone? Then:
Go back to the last commit.It’s back. Save point first, then courage. That’s the habit for every folder you care about: commit, let it try, look at what changed, keep it or go back.
2.4 What git is, and what it’s for
Git is the history of your folder. Every commit is a snapshot of every file as it was, with a date and a one-line message. It lives in the folder itself (a hidden .git folder), and nothing leaves your laptop unless you push it to GitHub (section 10). It was built in 2005 for the Linux kernel; today almost every software project uses it. You don’t need its commands: you ask, and the agent runs them.
What it gives you:
- Nerve. Let the agent try something bold. If you hate the result, go back.
- Seeing before accepting. “What changed?” shows every moved file and every edited line.
- A history you can ask about. “When did the sample drop from 240 to 220, and why?”
- Trying two versions side by side (a branch), without
analysis_v2_FINAL.R. - Sharing. The same history on GitHub, for co-authors and backup.
Six words worth knowing:
| Word | Means | You say |
|---|---|---|
| commit | a save point, with a message | “Commit this.” |
| diff | what changed since a save point | “What changed?” |
| log | the list of save points | “Show me the history.” |
| restore | go back, for one file or all | “Go back to the last commit.” |
| branch | a parallel version, to try something | “Make a branch to try X.” |
| push / pull | send to / fetch from GitHub | “Push.” / “Pull.” |
Commit before you let it loose, and again when it works. One thing git is bad at: it stores Word files, PDFs and spreadsheets, but it can’t show you what changed inside them. That’s one reason to keep notes and scripts in plain text (section 7).
3 What it knows about you
3.1 What it can see
The agent only knows what’s in its context window: your profile, the project file in this folder, this conversation, and the files it has opened. Nothing else. /clear empties the conversation and everything it read. Your profile and the project file load again.
Try it. Tell it one thing about yourself, for example what you teach. Type /clear. Then ask: “What do you know about me?”
3.2 Who is asking?
Without context, it answers the typical person who would type your question. “Can I deduct my home office?” gets US tax rules unless you say you work in Norway. A specialist is far from typical, so the default answer is aimed at someone else.
The demo from the room: “A: 48 purchases from 1,000 visitors. B: 62 from 1,000. Is B better?” With no profile, a cautious frequentist answer (p ≈ 0.17). With “I’m a Bayesian statistician; I read Gelman, Jaynes and McElreath”, a 91% chance B is better: ship it. With the single line “My intellectual idol is Jim Berger”, Claude reached for Sellke, Bayarri and Berger (2001) and said there’s at least a 45% chance of no real difference. One line changed the school.
You can’t opt out of being stereotyped. You can choose the stereotype.
3.3 Does it matter in your field?
In statistics, in the terminal, it barely matters. Fields with schools of thought may be different. Find out with your neighbour:
Is the value premium compensation for risk, or mispricing?Ask it once with no profile. Then add one line to the folder’s AGENTS.md: one of you writes “I’m in the Fama–French tradition”, the other “I work in behavioural finance; I read Shiller and Thaler.” Type /clear, ask again, and compare. Other questions that split schools: “Do active funds beat index funds after fees?”, “Is momentum a risk factor?”
3.4 Your profile: how you work
What reliably carries across every project is how you work: the rules you’re tired of repeating. For example:
- I work in R (or Stata, or Python).
- Cluster standard errors by firm unless I say otherwise.
- I write in LaTeX. Never give me Word unless I ask.
- Tell me what you’ll do before you change anything.
3.5 Two files
| About | Where | |
|---|---|---|
| Your profile | You: how you work, your field and school | .claude\CLAUDE.md or .codex\AGENTS.md in your home folder |
| Project file | This folder: the paper, the data, the rules | AGENTS.md in the folder (both tools) |
Your profile is read in every project; a project file only in its own folder. Both are plain text you can open in Notepad.
3.6 Exercise: write your profile
Two sentences on who you are: your field, your school, what you read.
Three rules you’d give any research assistant: your tools, your style, one pet peeve.
Ask:
Save these as my personal profile, the file you read in every project: [your sentences and rules]Claude asks permission first, because the file is outside this folder.
Test it in a fresh folder: ask for something your rules should change. Did they stick?
“Open my profile in Notepad.” (Mac: TextEdit.) That’s all it is.
Codex users: write more than one line. In testing, Codex ignored one-line profiles that Claude picked up.
Homework. Talk for 90 seconds with speech-to-text about how you work and who you are, then “Add this to my profile.” Keep adding: corrected it twice about the same thing? It goes in. In the chat apps, the “who you are” half matters most; put it in custom instructions.
4 Take a paper apart
Pick a recent paper, let the agent read all of it, find its biggest problem, and ask it to fix it. Finding a hole worth working on usually takes weeks of reading. Here it takes a minute, and you spend your time judging.
The floor: one problem you have checked against the paper yourself. A fix that holds up is the bonus.
4.1 Get a paper
Pick a recent paper in your field: one you know well (so you can judge the answers), or one you’ve been meaning to read. Make a folder for it and start the agent there:
mkdir paper
cd paper
claudeOpen PDF link (NBER, arXiv and similar):
Download this paper into this folder: [paste the link]SSRN or a journal website usually blocks the agent. Download the PDF in your browser, then:
Move the PDF I just downloaded into this folder.
If it says it can’t read the PDF, tell it to find a way. It may ask to install a small tool; say yes.
4.2 Ask
Talk to it like a colleague who has just read the paper.
What's the main claim here?Check this against your own reading. If it gets the claim wrong, everything after it is suspect.
What's the biggest methodological problem?Can you fix it?A theory paper or a case study works too: the problem may be an assumption, the sample, or the evidence behind the main result.
4.3 Push
The first answer is rarely the best one.
- “Show me where in the paper.” Every claim should come with a quote. Find it in the paper yourself; it takes seconds.
- “Do they address it anywhere, including the appendix?” A first critique is often already answered in the paper. This is the push people skip.
- “Is that really the biggest problem?”
- “That’s generic. What’s specific to this paper?”
- “Show that the fix works.” It may write and run a simulation, or fetch public data.
A fix takes 5–10 minutes: it writes and runs code. Let it work. If it asks to install Python, say yes.
Then: try it on your own working paper, before a referee does. Keep the folder: everything in the afternoon starts from it.
5 Referee panel
Who will referee your paper, and what will they say? Find out before you submit.
Real referees read your paper cold, alone, each with their own hobby-horses. A fresh session of the agent knows nothing but the paper and what you give it, so one fresh session per referee gives you independent reports. And names carry a lot: “as Campbell Harvey would” brings the factor zoo and the t > 3 bar with it.
The floor: two referee reports from fresh sessions, and one concern in them you hadn’t thought of.
5.1 Who would referee this?
In your paper folder:
Who are the five most likely referees for this paper, judging by its references? One line each: what would they care about?Argue with the list. Add the one you fear most.
5.2 Ask how it’s done
The most underused move with these tools: ask the agent how a job should be done before you ask it to do it.
I want simulated referee reports on this paper. How does a good finance referee work? How deep should they go? Should they try to reproduce the main results? Propose a procedure.Argue with it, then:
Save what we agreed as reviews/how-to-referee.md.5.3 One fresh session per referee
Type /clear, so the next referee starts cold, with only the paper and the procedure. Then:
Referee this paper as [name] would, following reviews/how-to-referee.md. Save it as reviews/[name].md.Repeat for each referee, with /clear before each.
In a hurry? Run them side by side: open a second PowerShell window in the same folder and start another claude or codex there. Claude Code can also do it in one prompt: “Run the referees as parallel subagents.”
5.4 Not what you wanted?
Change the procedure, not the report. Tell the agent, or edit reviews/how-to-referee.md in Notepad:
- Go deeper on identification.
- Try to reproduce Table 2.
- Check whether the data are available.
- One page. No summary of the paper.
Then /clear and rerun. If you’ll do this for every paper, make the procedure a skill (section 8):
Turn reviews/how-to-referee.md into a skill called referee-panel.5.5 The editor
/clear, then:
You're the editor. Read the reports in reviews/. Where do the referees agree, where do they disagree, and what must be fixed before submission?Then the useful bit:
Draft a plan for each must-fix concern: new analysis, rewriting, or an argument.5.6 Worth knowing
- Use it on your own paper, before a real referee does. A published paper is fine for practice.
- Different tools and different runs find different problems. Compare with your neighbour, or run Claude and Codex on the same paper.
6 References
Four jobs, one prompt each. Do them in any order, on your paper’s reference list or your own draft.
Why bother. An audit of 2.5 million biomedical papers found fabricated references in 1 in 277 papers in early 2026, up from 1 in 2,828 in 2023 (Topaz et al., The Lancet, 2026; a letter, and its method has been questioned). Closer to home: a 2026 paper in Structural Equation Modeling cites Foldnes & Grønneberg (2024), and the DOI resolves to a 2022 “Editorial acknowledgement”. The paper reports simulation results for methods from a source that doesn’t exist.
The models got better; the practice didn’t. A current model with search doesn’t invent DOIs. Its failures now are the wrong paper, or a paper it couldn’t fetch, and both take seconds to catch.
6.1 Search
Shallow and deep are two different requests, not two settings.
Shallow: one specific question. The trick is to be specific. Mine, in a field I know well:
What's similar to Hjort–Glad density estimation with a parametric start? Other kernel estimators with something that could count as a parametric start, also on domains other than the real line? Do any outperform it? Download the papers into literature/, and give me the DOI for any you can't get.Questions like this keep turning up important papers I didn’t know existed. Yours:
What's similar to [your idea, in one exact sentence]? Has anyone built something like it? Does it beat [the usual approach]? Download the papers into literature/, and give me the DOI for any you can't get.“Papers on [your topic]” gets you the famous ones you already know. A specific question gets you the one you missed.
Also quick: pinning down a paper you half-remember.
Find the actual paper: full citation, all authors, journal, year and DOI. [what you remember]. If you can't find it, say so. Don't guess.Deep: every paper. You’re finalizing, not exploring: the related-work section, a grant application, a referee response.
I need every method since [year] related to [your idea]: working papers too, and other names for the same idea. Make a table: paper, what it does, how it relates to mine. Take your time.It’s slow, and it uses a good part of your usage limit. Start it, then get coffee.
6.2 Keep the papers in the folder
The habit that makes everything else here work: every paper you need sits in the project folder, so the agent reads the paper itself instead of searching for it again, or relying on its memory of it. Make it a rule in the project file (section 8), once:
Add to the project file: papers go in literature/. Whenever we need one, download it there if you can get it. If you can't, give me the DOI so I can.Catching up on a paper you’ve already written:
Do that for every paper in my reference list. Put the DOIs you couldn't get in literature/to-fetch.md.- Paywalls stop it, and BI’s library takes the DOI list.
- SSRN blocks agents. Download in your browser, then “Move it into literature/.”
- Using git? “Keep literature/ out of git.” Track the text you wrote, not the PDFs you downloaded.
6.3 DOI check
Check every entry in the reference list. Does the DOI resolve to this exact paper: same title, authors and year? Make a table: OK, wrong paper, no DOI, not found. Fix what you can, and tell me how you know.Then check three yourself: paste each DOI after https://doi.org/. Same paper?
Practice file. A messy .bib handed over by a fictional co-author, with six planted traps (a DOI that points to the rebuttal, a sequel that was never written, and so on), plus two warm-up files. Ask the agent:
Download https://github.com/JonasMoss/vibing/raw/main/courses/2026-10-finance/props/references-practice.zip into this folder and unzip it. Then read messy-list.bib and do what the comment at the top says.6.4 Claim check
The one nobody does: does the cited paper say what it’s cited for? This is where the literature folder pays off: the papers are already there, so it quotes the real text instead of searching for it.
Take each citation in the introduction. Read the cited paper in literature/ and check whether it supports the sentence it is cited for. Quote the passage. Make a table: sentence, citation, supports / partly / doesn't / not in the folder, quote.- Run it on your own draft before you submit.
- “Not in the folder” is not “doesn’t exist”. Get those through the library, put them in literature/, and ask again.
A polished bibliography is not evidence of a correct bibliography.
7 Leave a trail
Turn a messy project into something a co-author could rerun, and check that they could. Practice on a messy one first, then do the same with a copy of one of your own folders.
The floor: the rerun test passes. After /clear, the agent reruns your folder from scratch and gets the same numbers.
7.1 The folder
A paper you’re writing with Kari Nordmann (fictional): Monthly or yearly? Price framing and willingness to pay. A survey experiment, cleaned by hand in Excel, analysed twice and written up in Word. Start the agent in an empty folder and ask:
Download https://github.com/JonasMoss/vibing/raw/main/courses/2026-10-finance/props/messy-project.zip into this folder and unzip it.What you get:
messy-project/
├── survey_export_2026-03-14.csv the raw survey, 240 rows
├── data_clean_FINAL.csv cleaned by hand in Excel
├── analysis.py first try
├── analysis_v2.py the analysis
├── table1.txt results
├── table1 (copy).txt results?
├── notes.txt "the crazy WTP ones?? ask Kari"
└── draft_v3.docx the Word draft
Could Kari rerun this? Could you, next spring?
7.2 Organise
First a save point, the same move as in section 2:
Set up git here and commit everything.Then:
Organise this folder so a co-author could rerun everything. Show me the plan first. Don't move anything yet.Argue with the plan: folder names, where the raw data goes, what happens to old versions. Nothing should be deleted; old versions can get a folder of their own. Then “Go ahead.”
Write a README for a co-author, and a project file for you. Then commit.The README is for people: what this is, and how to rerun it. The project file (AGENTS.md) is for the agent, which reads it every time it starts here.
7.3 The test
Type /clear. The agent now knows nothing about this folder, like a new co-author.
Rerun everything from scratch. Do you get the same numbers?Not the same, or it can’t? That’s the useful part:
What's missing? Fix it, so the next rerun works.Using the practice folder? Compare the rerun with the numbers in the Word draft. They don’t match, and finding out why is the point of the exercise.
7.4 Why text
| Invisible | Visible | |
|---|---|---|
| Writing | Word: you can’t see what changed | Text: every change shows |
| Data | Clicking in Excel: no record | A script: the steps are the record |
Writing in LaTeX? That’s text already. Writing in Word? Keep Word for the paper, keep the rest in text, and ask for a Word copy when you need one: “Make a Word copy of README.md.”
You may remember Reinhart and Rogoff’s “Growth in a Time of Debt” (2010): above 90% debt to GDP, average growth of −0.1%. In 2013 a graduate student, Thomas Herndon, couldn’t reproduce it, asked for the spreadsheet, and got +2.2%. Part of the gap was an AVERAGE over the wrong rows, which left out five countries; the rest was years left out and the weighting. Your own rerun finds that kind of thing first. A wrong script you can read beats a right click you can’t remember.
8 Notes and structure
Where things go, and how the agent keeps track of a project for you.
8.1 Four questions
Answer these about a real project, in your own words:
- What do you re-explain every single time? → that goes in the project file.
- Where does undone work live? (If the honest answer is “my head” or “my inbox”, that’s the finding.) → a to-do file.
- Where are you allowed to be wrong? → notes, drafts, scratch.
- What could you never rebuild? → raw data, the manuscript. Protect it.
Everyone already has a filing system. It just isn’t written down, so it can’t be delegated to a co-author or an agent.
8.2 An example: my paper folder
Every paper of mine starts like this (a skill sets it up):
my-paper/
├── AGENTS.md the rules, for the agent
├── README.md for people: what this is, how to rerun it
├── STATUS.md where we are, what's next
├── manuscript/ everything that ends up in the PDF
├── code/ everything that computes
├── results/ the numbers the paper relies on
├── dev/ notes, drafts, talks: in git, not published
└── local/ downloaded papers, raw runs: not in git
My four answers: the project file (AGENTS.md), STATUS.md, dev/, and manuscript/ with results/. Yours will look different, and that’s the point of the next step.
dev/ or local/? Both hold things that aren’t the paper. The difference is git:
| In git and on GitHub | Published with the paper | |
|---|---|---|
manuscript/, code/, results/ |
yes | yes |
dev/ |
yes | no |
local/ |
no | no |
dev/is work in progress: notes, drafts, talks. It’s committed, so it’s backed up on GitHub and your co-authors see it, but it never goes out with the paper.local/is listed in.gitignore, a plain-text file that tells git which files and folders to leave alone. Git never sees them, so they never reach GitHub: downloaded PDFs you may not redistribute (inlocal/literature/), big raw files, anything else that stays on your machine. To set it up: “Keep local/ out of git.”
8.3 Ask for a structure, not a template
Here is how I work: [your four answers]. Propose a folder structure for this paper and a project file that says where things go. Show me first; don't move anything.Argue with it. Then “Go ahead”, and “Commit” if you use git. Check nothing got lost: “How many files were there before, and how many now?”
8.4 Notes the agent keeps for you
At the end of each session, add what we did and what's next to notes/log.md. Put this rule in the project file.Keep literature/notes.md: one paragraph per paper we read, with the pages that matter.Next time you sit down:
Read the project file and the log, then tell me where we left off.8.5 The project file
A plain-text file in the folder that the agent reads at every start: what the project is, where things go, and your rules. Your profile is about you; the project file is about this folder.
One file: AGENTS.md. Codex and Claude Code both read it. One catch: if the folder also has a CLAUDE.md, Claude Code reads that instead, so keep just the one file. Both tools also read project files in the folders above where you start them, so start the agent in the project folder itself.
Getting one, and changing it. You don’t write it yourself; you ask:
Write an AGENTS.md for this folder: what the project is, where things go, and the rules you should follow. Show me first.Changes go the same way: “Add to AGENTS.md: never edit anything in data/raw.” Then read what it wrote: the first draft describes the folder, but only you know the rules. (/init drafts one too. In Claude Code it writes a CLAUDE.md, so ask for AGENTS.md by name instead.)
What goes in:
- what this is, in one paragraph: the paper, the question, the data;
- where things go: raw data in data/raw, papers in literature/, scripts in code/, the paper is LaTeX in paper/;
- the rules: never edit anything in data/raw; download papers into literature/, and give me the DOI when you can’t; only cite references with a verified DOI;
- how to rerun everything, in one command.
Keep out what it can see for itself (the file list), long essays, passwords, and anything about you (that’s your profile).
The habit that pays off. Corrected it twice about the same thing? Then:
Add that to the project file.Every correction you write down is one you never make again. Keep it short: it’s read at every start. It’s plain text, so open it in Notepad and fix it yourself whenever it’s wrong.
8.6 Skills, if you’re curious
Memory is the AI remembering you. A skill is you writing it down.
A skill is a saved workflow with a name and a note on when to use it. It’s worth making one for a routine you repeat, want done the same way every time, or want to hand to a colleague. You never touch the file format; you ask:
Turn what we just did into a skill called tidy-paper.From then on, “tidy this paper” runs the whole routine. Most things work fine without skills. Only bother for the ones you catch yourself repeating.
9 Your own work
Bring the real thing. Commit first (“Set up git here and commit everything”), then let it try.
A replication, with public data:
Get the Fama–French factors from Ken French's website and reproduce the summary statistics for RM−RF, SMB and HML in Table 2 of Fama and French (1993), July 1963 to December 1991. Then extend it to today: does it still hold?A project folder: “What’s here? What’s missing? What would you do first?”
A dataset: “Describe this data. What looks odd?”
A draft: “Where is the argument weakest?”
A referee report you received: “Draft a response plan, point by point.”
A course: “Turn my lecture notes into exercises with solutions.”
A lecture animation, like the compound-interest one from the morning:
Make a 15-second animated SVG that explains [compound interest] to first-year students. Three scenes, one caption each.It’s one file that plays in any browser. Change a number or a caption and ask again.
Anything you’ve been putting off because it’s tedious.
Before you leave: what will you try on Monday? Pick one thing, and put it in your calendar. And don’t overthink it: just use it.
9.1 Worth reading
- Epoch AI benchmarks: where to track what the models can do.
- One Useful Thing, Ethan Mollick: the best running account of using these tools for academic work.
- The Plain Person’s Guide to Plain Text Social Science, Kieran Healy. Written for social scientists, not programmers.
- Good enough practices in scientific computing, Wilson et al. (2017). For researchers who aren’t software engineers.
- Spreadsheet horror stories, EuSpRIG. Real incidents with price tags.
Questions afterwards: jonas.moss@bi.no
10 GitHub, if you want it
git is the save points on your laptop (section 2). GitHub is a copy of that folder, with all its save points, in the cloud. It survives a dead laptop, lets a co-author work on the same folder, and records every change with a date and a name. OneDrive keeps files in sync; GitHub keeps every version.
Why a researcher might want it: backup with history; co-authors on one folder instead of emailed attachments; replication packages (journals increasingly ask for code and data, and a GitHub repository archived on Zenodo gets a DOI); course pages built from plain text.
10.1 Set up, once
Make a free account at https://github.com. Private repositories are free.
Ask the agent:
Install the GitHub command-line tool and log me in.It installs
ghand opens your browser to log in.
10.2 Put a folder on GitHub
Put this folder on GitHub as a private repository called [name]. Check .gitignore first: no PDFs, no raw data, nothing confidential.Then, whenever you’ve committed something worth keeping:
Push my latest changes.Laptop dead? On the new one, ask: “Download my repository [name] into this folder.” Everything is back, with its whole history.
10.4 Pull requests
Instead of pushing straight into your project, a co-author can propose a change: a pull request. You read it, discuss it, and accept it (merge) or not. The discussion and the revisions stay attached to the change. It’s referee-and-revise, for a folder.
Their side:
Make a new branch, [the change], commit it, push it, and open a pull request to [owner] that explains what you changed and why.Your side:
Are there open pull requests on this repository? Summarise them.Check out the pull request and test it. Does everything still run? Do any numbers change?Then “Comment on the pull request: [your question]”, or “Merge it, then pull the latest changes.”
10.5 Exercise: in pairs
Both of you need a GitHub account and the GitHub tool (see “Set up, once”). Any folder works; the practice folder from section 2 already has its history.
A:
Put this folder on GitHub as a private repository called [name], and add [B's GitHub username] as a collaborator.B:
Accept the GitHub invitation from [A's username], and download [A's username]/[name] into this folder.(The invitation also arrives by email, if the agent can’t find it.)
B:
Make a new branch, [your change], commit it, push it, and open a pull request that explains what you changed and why.A: “Any open pull requests? Summarise them.” Then comment, or merge.
Swap roles.
10.6 Keep an eye on it
GitHub tells you. Click Watch on the repository’s page, and you get an email (or a push notification in the GitHub mobile app) for every new pull request.
Claude Code, while it’s open, can check on a timer:
/loop 10m check for new pull requests on this repository, and summarise any new ones.A reviewer on GitHub itself. Connect the Claude or Codex app to the repository once. Then comment
@claudeor@codex reviewon a pull request, and it posts a review there.
10.7 Public or private
- Private by default. Only you and the people you invite can see it.
- Public only for what you’d publish anyway: a replication package, a course page.
- History is forever. A file you delete later is still in the history, so never commit passwords, student data or confidential data in the first place.
- Research data: check the data agreement before it goes to any cloud service.
11 Connectors: email, calendar, files
MCP is a standard plug that lets the agent use other services: your email, calendar or files, the same way it uses your folder. Each plug is called a connector.
Switch one on (Claude). Go to claude.ai/customize/connectors, pick Gmail, Google Calendar or Google Drive, click Connect, and approve with your Google account. Start a new Claude Code session and type /mcp: it’s in the list. Codex supports MCP too; ask it how to connect yours.
Email:
Find all my emails from [name] this month. One line each: date, subject, what they want.Draft a reply. Show it to me before you send it.Calendar:
What does my week look like? Find two free hours for writing, and put them in my calendar.Drive:
Find the referee report on [paper] in my Google Drive, and draft a response plan, point by point.