Teacher notes
AI in the terminal · Department of Finance · Friday 16 October 2026
The shape of the day
Two goals, as always: they leave knowing these tools do far more than they thought, and believing they can do it themselves on Monday. Everything else is in service of that.
A spine and a menu.
- The morning spine is fixed, and all of it is mandatory. Everyone gets running with git, has the agent do something tedious on a messy folder (with a save point and an undo), writes a profile, and has the agent take their own paper apart. The “get started” goal is met by lunch, whatever happens after.
- The afternoon is a menu. Referee panel, References, Leave a trail, Notes and structure. Referee panel and References work on the paper folder from the morning; Leave a trail brings its own practice folder. Any subset in any order makes sense.
- There’s no pressure to cover the menu. The menu slide and the handout show the whole range. That is the range lesson, even for blocks we never open. Doing two modules well beats four in a hurry.
Where things are. One deck per module in decks/. Open decks/menu.html first: every card opens a module’s deck, and Menu in each deck’s footer comes back. Inside a deck, press M for the slide list and Esc for the overview. handout.html has every prompt with a copy button, section-numbered to match the slides.
| Deck | What | Minutes |
|---|---|---|
menu |
Title, the menu, “one folder”, start the install | 10 |
power |
Capability talk: programming, maths benchmarks, a released model on open problems, Navier–Stokes, the 6 October results, “the biggest day in history?”, Duminil-Copin; ends on a teaching slide (an animation written as code) | ~10 |
how |
What an LLM is (training: pre-training, feedback, RL; use: thinking), prompt engineering (RIP), what an agent is, usage limits | 6 |
setup |
Install, log in, git (required), which model and limits, first contact | 45 |
cleanup |
Do something tedious: the WISEflow export, with a git save point and an undo | 25 |
knows |
Context, who is asking, the Berger demo, does it matter in your field, how you work, your profile | 20 |
paper |
Take a paper apart; ends on the afternoon choice | 50 |
review |
Referee panel | ~40 |
refs |
Search (exercise), papers in the folder, DOI check, claim check | 40–60 |
trail |
Organise a folder, the rerun test, what the rerun found | ~50 |
notes |
Four questions, structure, notes, project file, skills | 20–30 |
own |
Open lab and close | 30+ |
agents |
If asked: the project file, one AGENTS.md for both tools, asking for changes |
10 |
github |
If asked: git vs GitHub, setup, co-authors, public or private | 10 |
mcp |
If asked: connectors (MCP), switching one on, one example each for email, calendar, Drive | 5 |
more |
Backup: security, benchmarks, Eco, Concord, prompting; the long capability talk | as needed |
Before Friday
- Links: done. All practice files are on GitHub
mainand tested. The handout’s download prompts (sections 2, 6, 7) use the GitHub links; the decks point to the handout. - TODO: go through Leave a trail in more detail. It overlaps with the morning’s cleanup: both organise a messy folder with a git save point. One option: refocus the exercise on reproducing Table 1 from the raw survey export, so the planted problem surfaces there.
- Course files.
course-files.zip(handout plus the three practice zips) is what people download from the course page, vibing.jonasmoss.com → Finance (on the menu deck’s “Get the course files” slide). The page also has the handout and every deck. After any change to the handout or a prop: re-render, thenpython3 make-course-files.py, commit, push. Bring the zip on a USB stick too. - Test on a BI-managed Windows laptop, on campus wifi: install, log in, one prompt, one PDF read, and git without admin rights (the portable route). Also check that
/usage(Claude Code) and/status(Codex) still show the plan limits in Friday’s versions. Dagfinn’s Codex install stalled in August until he picked the restricted (unelevated) mode. Codex reading PDFs on bare Windows is still untested.- Tested 14 October on a BI PC: the Claude Code install needed no admin rights, but asked for
.local\binto be added to PATH by hand. Fixed: the Claude install is now two lines, the second adds it to the user PATH. - Not tested on a BI PC: right-click → Open in Terminal,
explorer ., and whether Documents sits in OneDrive. If it does, keep practice folders in the home folder (wheremkdir practiceputs them): git and OneDrive syncing don’t mix well. Demo all three on your own Windows laptop live.
- Tested 14 October on a BI PC: the Claude Code install needed no admin rights, but asked for
- Demo papers: chosen.
- “Take a paper apart”: your own pick.
- Referee panel: Nagel, “Seemingly Virtuous Complexity in Return Prediction”, NBER w34104 (https://www.nber.org/papers/w34104; NBER downloads directly). The pilot answers are in
scratch/examples/paper-critique/pilot/, local only. - Replication card: Fama and French (1993), Table 2 (see “Your own work” below).
- The Foldnes example. Make only the airtight claim: the paper they cite does not exist; the DOI resolves to a 2022 editorial acknowledgement. Don’t say “the method doesn’t exist” unless you’ve settled it with Foldnes. Details are in
scratch/examples/references/README.md. - Brief Ludvik. In the morning he roams the install, Windows first. In the afternoon he’s a second help desk. Tell him the rule: two minutes stuck, hand up.
- Room. A2 Blue 1, booked 09:00–16:45. Check that power reaches every seat; bring extension cords.
- Pre-run the backups on Claude: the Berger demo (one-line profile), plus one referee report from a fresh session.
Run sheet
| Time | Block | Minutes | Notes |
|---|---|---|---|
| 09:00 | Welcome, menu, “one folder” | 5 | Show the menu: this is the whole range. |
| 09:05 | Start the install | 5 | Everyone pastes the lines for their tool and leaves the lid open. Downloads run during the talk. |
| 09:10 | Capability talk | ~10 | power. Ends on the animated teaching slide (let it play once). more has the long version, for questions. |
| 09:20 | How it works | 5 | how. Don’t elaborate. “Prompt engineering (RIP)” follows the thinking stage: the 2022 trick is now built in. The limits slide sets up “why did it stop?” before it happens. |
| 09:25 | Get it running | 45 | setup. Install, log in, git for everyone, first contact. |
| 10:10 | Do something tedious | 25 | cleanup. Save point, reorganise, what changed, the undo, then four slides on what git is. |
| 10:35 | Break | 10 | Stragglers keep installing, with Ludvik. |
| 10:45 | What it knows about you | 20 | knows. Berger demo, describe yourself. |
| 11:05 | Take a paper apart | 50 | paper. Demo 10, hands-on 30, round-up 10. |
| 11:55 | Choose the afternoon | 5 | Show of hands on the “After lunch” slide. |
| 12:00 | Lunch | 60 | |
| 13:00 | Menu block 1 | ~40 | |
| 13:40 | Menu block 2 | ~40 | |
| 14:20 | Break | 15 | |
| 14:35 | Menu block 3, or own work | ~45 | |
| 15:20 | Your own work | 30 | Open lab. Walk the room. |
| 15:50 | Before you go | 10 | “What will you try on Monday?” Collect a few answers aloud. |
When behind: cut talk, never hands-on time. The morning is tight and has no buffer. If “Take a paper apart” runs over, let it run into the start of the afternoon and drop one menu block. Don’t cut the cleanup or the profile: they’re the spine.
If the room can’t decide on the afternoon, take this default:
- References, with the claim check as the centre. Everyone has a reference list, and catching one bad citation in your own field lands hard.
- Referee panel. The most “I didn’t know it could do that” of the four.
- Leave a trail, if people are still keen, or straight into own work.
Notes and structure is the lightest module. It folds naturally into the end of Leave a trail.
Module cards
Get it running
- Aim: a blinking cursor that answers, and
git --versionthat answers, on every laptop. - Floor: make a file, read it back, delete it.
- Git is required now. If someone has no admin rights, the agent installs the portable Git for Windows into their user folder. Untested on a BI-managed laptop: try it on Thursday. If it fails on the day, they pair with a neighbour for the git steps.
- Watch for:
- the “x86” PowerShell;
- skipping “close and reopen”;
- Command Prompt instead of PowerShell;
- Codex’s admin or sandbox prompt (choose unelevated);
- git needing admin rights (the portable route, above).
- People with no subscription pair up. Don’t let them sit idle.
- Finished first? They’re the help desk. Say it out loud.
Do something tedious
About 25 minutes. The prop is props/decrap-me.zip, a synthetic WISEflow export: 15 group submissions with junk file names, cover sheets, and an empty grade sheet in submission order.
- Aim: a reorganised folder, a grading table, and one deliberate undo.
- The framing: a job the chat window can’t do, because it can’t touch your files. Tedious, local, easy to judge.
- The git beat, from the August plan:
- Brief the task, but don’t start it.
- Ask the room: “Before you hit enter, does anything about this make you nervous?” Let someone say “what if it ruins something?”
- Answer with “Set up git here and commit everything.” Then: “Now it can’t do anything you can’t undo.” The answer lands because the question is live.
- The undo is on purpose: delete the Marking guidance folder, then “go back to the last commit”. Watching it come back is what makes git stick.
- Then four git slides (about 5 minutes): what git is (a timeline diagram; click for “go back”), what it’s for, branches and GitHub (a diagram; two clicks), and six words worth knowing. They come after the undo on purpose: people have felt it work, now they want to know what it is. Don’t go deeper than the slides. Branches and push/pull come back in the GitHub deck if anyone asks.
- Ties forward: “Leave a trail” and the GitHub demo reuse the same save-point move.
What it knows about you
About 20 minutes. Two halves, framed honestly.
- “Who you are” is an open question for this room. In stats, in the terminal, you barely notice it; fields with schools may differ. Say exactly that, and let the room find out. Pairs pick opposite schools on “Is the value premium compensation for risk, or mispricing?” (Fama–French vs behavioural). The round-up is data: did it change, and how much?
- Untested. Try the value-premium question once before Friday, with both one-liners, so you know roughly what to expect.
- The Berger demo is the tested example of a school changing the answer. The A/B question with no profile, then one line in the folder’s
AGENTS.md: “My intellectual idol is Jim Berger.”/clear, ask again. (Tested withCLAUDE.md; Claude readsAGENTS.mdthe same way as long as there’s noCLAUDE.mdin the folder or above it, so demo in a clean folder.)- Demo on Claude: Codex ignored the one-liner in testing.
- The backup result is on the slide after the demo.
- If a statistician objects that 0.91 is just 1 minus the one-sided p: they’re right; under a flat prior it’s the same information. The point is that the headline, the framing and the recommendation flip.
- “How you work” is the reliable win. The slide shows two lines from your own profile files (the 4-core rule, MathML). Your terminal profiles hold no “who I am” at all, only working rules, and you can say so.
- The “who you are” profile belongs in the chat apps. The homework slide says so, with your chat profile as the example. The coffee table (“Four askers”) is in
more.html#/chatif someone wants the chat-app version.
Take a paper apart
- Your demo: a paper you can judge. Ask the three questions, and push at least once on camera.
- Say these two sentences, no slide:
- when the first answers come back: “It turned the PDF into text before it could read it”;
- when the fix runs: “That script is the record of what it did.”
- Snags:
- SSRN returns 403 to the agent: download in the browser, then “move it here”.
- The fix needs Python; let it install.
- Codex on bare Windows may need a PDF tool; say yes.
- Watch for:
- People skip “Do they address it anywhere?” and report false alarms. Collect one for the round-up.
- Generic first answers are normal; “what’s specific to this paper?” fixes most.
- Honest framing: “I can’t tell you whether a critique of a finance paper is right. You can.” If it fails on someone’s paper, say so in the round-up. That’s a result too.
- From the September pilots on two NBER finance papers:
- the critique took 1–1.5 minutes, the fix 5–7.5;
- the two tools picked different “biggest problems” in the same paper.
Referee panel
New module, not yet run with a room. About 40 minutes.
- Aim: two or three independent referee reports from one agreed procedure, plus an editor’s letter.
- Floor: two reports from fresh sessions, and one concern they hadn’t thought of.
- The flow (the diagram on the first slide):
- Who would referee this?
- Ask how it’s done: a procedure, saved to
reviews/how-to-referee.md. /clear, then one referee at a time, following the procedure.- Not right? Change the procedure, not the report.
- The editor.
- The lesson to sell beyond referees: ask the agent how a job should be done before you ask it to do it. It’s the most underused move. Step 2 is the place to say so out loud.
- Why
/clearand not a new window: it’s the same in both tools and needs no window juggling. After the procedure discussion, it’s exactly right: each referee starts with only the paper and the procedure.- Parallel sessions are a one-liner for the eager: a second PowerShell window with its own
claude/codex, or Claude Code’s parallel subagents. - The agent can also start
claude -porcodex execin the background itself. Don’t teach that today.
- Parallel sessions are a one-liner for the eager: a second PowerShell window with its own
- Skill hook: “Turn the procedure into a skill called referee-panel” connects to the skills slide in
notes, if anyone asks. - Demo: the first two steps on your finance paper, live. Then one referee from a fresh session.
References
About 40 minutes for two of the four parts; 60 for all of them. Open with the “in the literature right now” slide (3 minutes), then let the room pick parts.
- Search. Tell the story first: shallow, specific questions have turned up important papers you didn’t know, in your own field (the Hjort–Glad prompt). Then the exercise, about 10 minutes: everyone asks one specific question about their own idea, and has it download the papers. The lesson is be specific. Deep (“I need every method since 1995…”) is a different request: finalizing, slow, and expensive on usage. Show the prompt; don’t run it in the room.
- Papers in the folder. The practice to sell hardest in this module: one rule in the project file (papers go in literature/; download what you can, a DOI for the rest), and the agent stops searching for the same paper again and again. It’s also what makes the claim check work.
- DOI check. The practice
.bibhas six planted traps. The full key is inscratch/examples/references/README.md. Floor: three DOIs pasted into doi.org by hand. - Claim check. The centrepiece. Run it on their own draft if they have one, after the papers are in literature/. “Not in the folder” is not “doesn’t exist”.
- The line to end on: a polished bibliography is not evidence of a correct bibliography.
Leave a trail
About 50 minutes. The full instructor key is in scratch/examples/leave-a-trail/README.md.
One story: Reinhart and Rogoff. Economists remember it, so ask the room first: “Who remembers the 90% debt threshold?” Keep it about their own interest, not open science: you want your rerun to find it before a grad student does. The facts, to stay accurate: above 90% debt to GDP they reported −0.1% average growth; Herndon, Ash and Pollin (2013 working paper; Cambridge Journal of Economics 2014) got +2.2%. The Excel range error (rows 30–44 instead of 30–49, dropping Australia, Austria, Belgium, Canada, Denmark) accounts for only about 0.3 points of that; the rest is excluded years and the equal weighting by country. Don’t say the Excel error alone flipped the result.
The folder is a practice project, not their paper: Kari Nordmann’s and their survey experiment on price framing (fictional), in
props/messy-project.zip. The deck shows the link and the file tree. Those who finish go on to a copy of a real folder of their own, or the morning’s paper folder.Floor: the rerun test passes after
/clear.Opening: a show of hands, LaTeX or Word? LaTeX users already write in text; for them the news is git and the rerun test.
The practice folder (
props/messy-project.zip) has a planted problem. Three willingness-to-pay outliers were deleted by hand in Excel, with no record. With them in, the effect flips sign and stops being significant. Only the rerun test finds it.The last-but-one slide is the reveal (“What the rerun found”: N = 223, −292 NOK, p = 0.26 with the three rows; N = 220, +53 NOK, p = 0.02 without). Show it only after the room has run the test. The handout only hints at it.
Git was set up in the morning, so the exercise’s first step is one prompt. Anyone who skipped the morning: the agent asks for a name and email at the first commit.
The test slide is a diagram with two clicks: “Same numbers, done”, then the loop back (“Not the same? Fix the README, run the test again”).
Not yet piloted with an agent end to end.
Notes and structure
About 20–30 minutes; lighter than the rest.
- The four questions come from the
tidy-projectskill in this repo (.claude/skills/tidy-project/SKILL.md). It’s a good live demo if someone asks “what does a real skill look like?” - “My paper folder” is your
paper-initlayout (five folders: manuscript, code, results, dev, local), simplified: no justfile, Modal or Slurm. Say it’s a skill, which sets up the skills slide at the end. - “dev/ or local/?” is the one distinction worth slowing down for:
dev/is committed (backed up on GitHub, co-authors see it) but never published;local/is in.gitignore, so git and GitHub never see it. It’s also the natural place to say what.gitignoreis. - Question 3 (“where are you allowed to be wrong?”) is the one people stall on, and the stall is informative.
- The log habit (
notes/log.md, then “tell me where we left off”) is the most useful thing in this module for people who’ll use the tool weekly.
Skills, if asked
Not taught. One slide at the end of the notes deck (notes.html#/skills) and a box in the handout.
- The line: memory is the AI remembering you; a skill is you writing it down.
- When it’s worth it: a routine you repeat, want identical every time, or want to hand to a colleague.
- Live demo if pushed: open the
tidy-projectskill and show it’s a page of plain English. Or, after any module: “Turn what we just did into a skill called X”, and show the file it wrote.
The project file, if asked
- The deck:
agents.html, about 10 minutes. The short version is already the last content slide ofnotes, and “Two files” inknows. - The practical answer people want: one file,
AGENTS.md. Codex always read it; Claude Code reads it since v2.1.277 (18 September 2026), but only when there’s noCLAUDE.mdin the folder or above it. An older repo with a one-lineCLAUDE.mdsaying@AGENTS.md(like this course repo) still works. - Changes go through the agent: “Add to AGENTS.md: …”. Nobody needs to open the file in an editor, though they can.
- The habit worth selling: “Add that to the project file” after the second identical correction.
/init: fine as a shortcut, but the draft only describes the folder. The rules are the valuable part, and only they can write those. Claude Code’s/initwrites aCLAUDE.md, which would then hide theAGENTS.md: ask forAGENTS.mdby name instead.- The gotcha: both tools also read project files from the folders above. Starting the agent in Documents with a stray
AGENTS.mdthere affects everything below it.
GitHub, if asked
- The deck:
github.html, about 10 minutes; section 10 of the handout. - The one-liner: git is save points on your laptop, GitHub is the copy in the cloud. OneDrive syncs files; GitHub keeps every version and who made it.
- Setup is one prompt once they have an account: “Install the GitHub command-line tool and log me in.” On Windows the agent runs
winget install --id GitHub.cli -e, thengh auth login, which opens the browser. Not tested on a BI-managed laptop; if winget is blocked, the installer is at https://cli.github.com. - The selling point for finance: replication packages. A repository archived on Zenodo gets a DOI.
- Say once: private by default; history is forever, so data and passwords never go in.
The pull-request demo, with Ludvik
About 10 minutes, live. It shows what collaboration looks like when neither of you types a git command.
Done live, from scratch. No pre-made repository. If something breaks, narrate the fix: watching the agent recover is part of the lesson.
Before Friday: Ludvik has a GitHub account, and gh installed and logged in. Have his username ready.
Setup, live (2–3 minutes): in a copy of the “Leave a trail” practice folder (props/messy-project.zip): “Set up git, commit everything, put this folder on GitHub as a private repository called finance-demo, and add [Ludvik’s username] as a collaborator.” Ludvik accepts the invitation (it comes by email), then: “Download JonasMoss/finance-demo into this folder.”
The pull request:
- You, first: start the watcher in Claude Code and let it sit:
/loop 1m check for new pull requests on this repository; when one appears, summarise it.Test/loopbeforehand. If it misbehaves, just ask “any new pull requests?” when Ludvik says he’s done. - Ludvik, on the projector or his own screen, dictates or types: “Make a new branch, add the missing rule to the cleaning (drop willingness to pay above 5,000 NOK; it was done by hand in Excel and never written down), rerun the analysis, commit, push, and open a pull request to Jonas explaining what you changed and why.” The practice folder has no cleaning script, so his agent writes one. That’s fine, and it’s a bigger change to review.
- You: the watcher reports it. Then: “Check out the pull request and test it. Rerun everything from the raw export: do we get table1 (N = 220)? And what happens without the new rule?”
- With the rule: N = 220, +53 NOK, p ≈ 0.02. The hand-cleaned file is now reproducible from raw.
- Without it: N = 223, the effect flips sign (about −292 NOK) and stops being significant.
- Ludvik: “Address Jonas’s comment and push.” You: “Merge it.”
- Show the pull request page in the browser: the description, the comment, the commits, the merge. “That page is the whole history of the decision.”
Then the pairs exercise (deck slide “Exercise: in pairs”, handout section 10), about 15 minutes: A puts a folder on GitHub and adds B; B accepts, downloads, and opens a pull request; A reviews and merges; swap. Everyone needs a GitHub account and gh logged in first, which is the slow part: start that during the demo. Accepting the invitation through the agent goes through the GitHub API (gh api user/repository_invitations); if the agent fumbles, the invitation email has a button.
Optional, if set up beforehand: comment @claude (or @codex review) on the pull request, and let the room watch a review appear on GitHub. This needs the Claude GitHub app (/install-github-app in Claude Code) or Codex’s GitHub connection on the demo repository. Not yet tested.
Your own work
- Walk the room. Ask “what are you trying?”, not “is it working?”
- Stuck people: suggest “What’s here? What’s missing? What would you do first?” on a real folder, after a commit.
- The replication card is the one to push for this room: Ken French’s data library is public and agents can download from it, so a classic table plus “does it still hold today?” is doable in 20 minutes. The target: Fama and French (1993), JFE 33, Table 2, July 1963 to December 1991, 342 months. Published monthly means: RM−RF 0.43% (sd 4.54, t = 1.76), SMB 0.27% (2.89, t = 1.73), HML 0.40% (2.54, t = 2.91). Today’s library won’t match to the decimal (CRSP has been revised since, and the factor construction updated): close but not exact is the expected result, and “why not exact?” is a good question to ask the agent. “Does it still hold?” leads straight into the value-premium debate from
knows.- Checked 14 October (library file built from the August 2026 CRSP), monthly %, mean (t): same sample, RM−RF 0.41 (1.65), SMB 0.26 (1.67), HML 0.38 (2.75): within 0.02 of the paper.
- After the paper, January 1992 to August 2026 (416 months): RM−RF 0.76 (3.56), SMB 0.06 (0.36), HML 0.23 (1.40). The size premium is gone and value has halved, while the market did better. That’s the punchline people find themselves.
- Close: “What will you try on Monday?” Get three answers out loud. Remind them the handout has everything, including what we skipped.
Questions you’ll get
The longer prepared answers are in scratch/august-course/grievance-answers.md. The short versions:
- “Is my data safe?” What it reads is sent to the company, like any cloud upload. Keep out personal data and confidential material.
- “Can it read my email?” Yes, through connectors (MCP): Gmail, Google Calendar and Drive connect in a few clicks at
claude.ai/customize/connectors, and Claude Code picks them up (/mcplists them). Deck:mcp.html, five slides. BI mail is Outlook: a Microsoft 365 connector exists, but BI IT would likely have to approve it (not checked). Don’t demo on your own work mail. - “Does it train on my data?” It depends on the plan and changes over time, so check your plan’s terms. When unsure, assume it could be retained.
- “Claude or ChatGPT?” Both are excellent. Use the one you pay for. They aren’t interchangeable, so trying both on the same task is instructive.
- “Will the department pay?” That’s a question for Dagfinn. Marketing reimburses its people via its administration.
- “Can I get the slides?” Yes: each deck is one file. Zip
decks/*.htmland send them to Dagfinn afterwards with the handout. Openmenu.htmlfirst.