Teacher notes
AI in the terminal · Department of Marketing · 3 December 2026 · work in progress
The shape of the afternoon
Two goals, as always: they leave knowing these tools do far more than they thought, and believing they can do it themselves on Monday.
- One afternoon, one paper. After setup, the cleanup and a short capability talk, everything runs on one paper: theirs, or
ai-label-paperfrom the course files. At Strategy, people sat idle and didn’t see the point of the exercises (debrief, note 2). On their own paper, every exercise has an obvious point, and there is always a next thing to do. - The app, not the terminal. Strategy struggled with the terminal and it coloured the rest (debrief, note 1). The Claude app’s Code tab is the same agent with a window around it; ChatGPT users get Codex in the ChatGPT app.
- Start a job, talk while it runs, then collect. Take it apart runs while you give the sycophancy talk; the three referees run while you give What it can touch. The talks cost no hands-on time, and nobody waits for an agent (debrief, note 3).
- The referee panel is the centrepiece (the most engaged block at Strategy), now blind, and with sycophancy said out loud (debrief, note 5).
- Research only. Francesco: “4 hours methodology only”. No teaching block. References and DOI checks are in the handout, not the main line (debrief, note 4).
- No parking lot (debrief, note 6). Answer at the laptop, or at the very next break, the same day.
| Deck | What | Minutes |
|---|---|---|
menu |
Title, the menu, the course files | 5 |
setup |
The app, sign in, a folder, git by asking, the row under the prompt box, voice | 25 |
power |
Capability talk | ~10 |
cleanup |
Do something tedious: the WISEflow export, a save point, an undo | 25 |
yourpaper |
Which paper, training off, open it, a save point | 15 |
paper |
Take it apart: a mode for working alone, start it, leave it; then push | 5 + 20 |
sycophancy |
It wants to please you: three forms, why, the evidence, our measurement | 15 |
review |
Blind referee panel: the package, the brief, three referees, the editor, the comparison | 20 + 25 |
access |
What it can touch, while the referees run | 20 |
own, refs, git, words, more, … |
Any time; the handout has them all |
Before the day
- Room: A2 Blue 9 (Christine’s invitation). Power at every seat; bring extension cords.
- Reminder email, a week before (via Christine): bring a paper in progress, as close to submission as you have: a PDF of the draft, plus the Word or LaTeX file and the data and code if you can share them. No paper? We have one. And: install the Claude or ChatGPT app beforehand if you can.
- Test on a BI-managed Windows laptop: the list in the README (“To test on a Windows laptop”). Above all: does the Claude app install without admin rights, and can the agent install git in the user folder?
- Pre-run on
ai-label-paper: one Take it apart and one blind panel, and keep the outputs as backups for the projector. The key to what’s planted:scratch/examples/ai-label-paper/KEY.md. - Bring: sticky notes in two colours (two per person), extension cords, the course zip on a USB stick.
- Course files:
course-files.zip(the handout and the practice folders, includingai-label-paper). After any change to the handout or a prop: re-render,python3 make-course-files.py, commit, push.
Derailments
The rule: the room’s time is for the room.
- One-laptop questions are answered at the laptop.
- Good questions at the wrong time: one sentence now, or “at the break”, and then really at the break. (At Strategy, parked questions were never answered.)
- Privacy and access: the privacy slide comes before anyone opens their paper, and
accessruns while the referees work. - Level differences: fast people get the next step on their own paper. There is always one.
Run sheet: Thursday 3 December, 12:00–16:00
| Time | Block | Minutes | Notes |
|---|---|---|---|
| 11:30 | Install drop-in | 30 | Anyone can come early. |
| 12:00 | Welcome, menu | 5 | “One afternoon, one paper.” Sticky notes: yellow fine, pink stuck. |
| 12:05 | Get it running | 25 | setup. At 12:25, walk the rows: an app that answers, and git. |
| 12:30 | Capability talk | 10 | power. Stragglers keep installing. |
| 12:40 | Do something tedious | 25 | cleanup: a save point, the cleanup, an undo. |
| 13:05 | Your paper | 15 | yourpaper: training off, open it, “what is this?”, a save point. |
| 13:20 | Start: Take it apart | 5 | paper: the mode, the prompt, leave it. |
| 13:25 | Talk: It wants to please you | 15 | sycophancy, while it runs. |
| 13:40 | Break | 10 | |
| 13:50 | Back to Take it apart | 20 | Read take-apart.md, push, round-up. |
| 14:10 | Blind panel: package and brief | 20 | review 1–2. Check the PDF properties. |
| 14:30 | Start three referees · talk: What it can touch | 20 | review 3, then access while they run. |
| 14:50 | Break | 10 | |
| 15:00 | The editor, the comparison | 25 | review 4–6. “Did any referee know who wrote it?” |
| 15:25 | Keep it | 20 | The skill; own work; the handout for the rest. |
| 15:45 | Before you go | 15 | “What will you try on Monday?” Three answers aloud. |
When behind: cut talk, never hands-on time. power can shrink to five minutes; access can shrink to its stance slide. Never cut the save point before Take it apart.
Module cards
Get it running
About 25 minutes, first thing.
- Aim: by 12:25, every laptop has the app open in a folder, answering, and git installed.
- Floor: make a file, read it back, delete it, with the folder open in File Explorer so they watch it happen.
- New for Marketing: the app, not the terminal (Strategy debrief, note 1). It’s the same agent: the Code tab is Claude Code, and the ChatGPT app has Codex. No PowerShell, no PATH, no “close and reopen”.
- Not tested on a BI-managed laptop yet: the README’s list. Above all: the install without admin rights, and the agent installing git in the user folder. The deck and handout are written from Anthropic’s desktop docs (code.claude.com/docs/en/desktop-quickstart, checked 24 October): the Code tab needs Pro, Max, Team or Enterprise; Local → Open folder…; Ctrl+N for a new session; modes Manual, Accept edits, Plan, Auto. On native Windows, git is optional for the agent itself (it uses PowerShell without it), but we need it for save points.
- Fallback, if the app won’t install: the terminal route at the end of handout section 1. The Claude Code install was tested on a BI PC on 14 October: no admin rights needed, and the second line puts it on the PATH.
- Watch for:
- a free-plan account (the Code tab asks them to upgrade);
- Documents synced to OneDrive (use the home folder instead);
- a session in the wrong folder (the folder control under the prompt box);
- an old app without the folder control (update it);
- ChatGPT users: where Codex sits in the ChatGPT app is unchecked. Ask them to show you; note it for next time.
- People with no subscription pair up. The department reimburses one (receipt to Christine); say so.
- Finished first? They’re the help desk. Or the handout’s customer-journey web page (untested).
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. Kept as the first hands-on task (Strategy debrief).
- Names only. Group and participant ID are in every folder name, so the job needs no reading. In Finance the agent opened the PDFs anyway and the room got worked up. The prompts say “Work from the names only; don’t open the submissions”, and the last slide, “You decide what it reads”, makes it the point.
- The git beat: this is their first save point of the day. Brief the task, then ask “Before you hit enter, does anything about this make you nervous?” Answer with “Set up git here and commit everything.”
- The undo is on purpose: delete the Marking guidance folder, then “revert to the last commit”, with File Explorer open. Watching it come back is what makes git stick, and they need the habit for their paper twenty minutes later.
Your paper
About 15 minutes. The pivot of the afternoon: from here on, everything runs on one paper.
- Aim: everyone has a paper open in a session, described back correctly, with a save point.
- The privacy slide comes first, and fast: training off, leave out what you can’t share, tell co-authors. Then move on. The switch labels (Help improve Claude, Improve the model for everyone) weren’t verified for December: check your own account before the day.
- Who uses
ai-label-paper: anyone without a paper, and anyone whose data they can’t share. A good hybrid: their own PDF for the panel, our folder for Take it apart. Say it out loud; it removes the excuse to sit idle. - Watch for:
- a Word-only draft: whether the app reads
.docxwithout Python is untested. If it struggles, “save it as PDF”, or let it install a converter; - the paper folder inside OneDrive (copy it out);
- huge data folders (copy only what the paper needs).
- a Word-only draft: whether the app reads
- Done early: the
AGENTS.mdprompt. It pays off in every later session.
Take it apart
5 minutes to start, then the sycophancy talk, a break, and 20 minutes to collect.
- New for Marketing: start it and leave it (Strategy debrief, note 3). At Strategy this was the block that landed, but the room waited for it. The prompt writes everything to
take-apart.md, so the result survives the wait and the break. - The mode matters: Accept edits or Auto before they leave it, or it sits waiting for a “yes”. Say it twice.
- On
ai-label-paper(the key:scratch/examples/ai-label-paper/KEY.md), a good run finds:- the exclusions beyond the preregistration: with only the preregistered one, the candle effect goes from p = .030 to p = .166;
- the draft’s numbers come from an older export (the draft says N = 476; the script now gives 545);
- moderation claimed from simple effects, with no interaction test (p = .346 in the draft’s own data);
- the bonus, and the good news:
auth_3is reverse-worded but not reversed; reversed, α goes from .56 to .84.
- Not yet tested end to end with this prompt. Pre-run it on
ai-label-paperbefore the day (Before the day, item 4), and keep the output for the projector. - Snags: the script is R. If R isn’t installed, the agent translates it to Python or asks to install R; either is fine.
- The round-up question that matters: “Did it find something good for your paper?” It turns the exercise from fault-finding into help.
- Honest framing: “I can’t tell you whether a critique of a marketing paper is right. You can.” A false alarm is a result too.
It wants to please you
About 15 minutes, while Take it apart runs. A talk, no exercise.
- The order is the argument: three forms; why; the referees are worth having (capability before caveat); status (21%); the companies train against it; our measurement; what we learned; what to do. Ends on the panel.
- Our measurement, said accurately: two small models (Claude Haiku, GPT-6-Luna), one fictional paper with a clear flaw, fresh sessions without profile or memory, 24 October.
- Asked for a score under four framings: no difference (means 3.0–3.2 out of 10). Pushback, wrongly: 20 of 20 kept the concern.
- Asked for feedback on “my paper, I’m proud of it”: the verdict disappears (Haiku 10 → 1 of 10, Luna 9 → 0) and praise comes first (0 → 8, 0 → 10). A blind judge (GPT-6-Luna, same family as one writer) rated Luna’s feedback 1.9 vs. 4.0 out of 10. All 40 reports still asked for the interaction test.
- Scripts and aggregates:
scratch/examples/ai-label-paper/sycophancy/; raw reports inprivate/sycophancy-runs/.
- Don’t oversell either way. Not “they flatter”, not “it’s solved”. The line: asked as a referee, for a verdict, they stayed honest; asked “what do you think of my paper?”, they coached.
- Sources (checked against the originals, 24 October 2026):
- Liang et al. (2023), arXiv:2310.01783: overlap 30.85% vs. 28.58% (Nature journals), 39.23% vs. 35.25% (ICLR); 57.4% found the feedback helpful; 82.4% more beneficial than at least some human reviewers.
- Vasu et al. (2025), arXiv:2509.13400, Table 2: QwQ-32B, 21.4% of rejects become accepts with a stronger affiliation; 7.9% of accepts become rejects with a weaker one. Other models differ (DeepSeek R1 8B: 19.05% and 23.02%). 252 ICLR 2025 papers.
- OpenAI, GPT-5 System Card (August 2025), section 3.3, Table 4: 0.145 (GPT-4o), 0.052 (gpt-5-main), 0.040 (gpt-5-thinking); online prevalence −69% (free) and −75% (paid) vs. GPT-4o.
- Sharma et al. (2023), arXiv:2310.13548: the abstract’s “both humans and preference models prefer convincingly-written sycophantic responses over correct ones a non-negligible fraction of the time”; section 3.2, Claude 1.3 wrongly admits mistakes on 98% of questions.
- The GPT-4o rollback dates (rollout 24–25 April 2025, rollback from 28 April) are from Codex’s reading of OpenAI’s two posts (29 April and 2 May); not re-checked by hand. Codex’s full notes:
private/research/sycophancy-evidence.md.
Blind referee panel
20 minutes (package and brief) before the break, 25 (editor and comparison) after. The centrepiece: at Strategy, the referee panel was the most engaged block of the two days.
- Aim: three reports from fresh sessions in
review-room, an editor’s letter, and the comparison with “my paper”. - Floor: two reports, and one concern they hadn’t thought of.
- Why a separate folder: a session reads the project files in its folder and the folders above it.
review-roomsits next to the paper folder, so the referees never see the paper’s folder, data orAGENTS.md. - What else can leak (tested 24 October, in the terminal, with Haiku in a neutral folder): a fresh session sees the user’s profile in full, the home-folder path (with the user name), the account e-mail, and the list of the user’s skills. With the profile switched off, the e-mail and the path remain. So the paper itself must carry no names (the package step), the brief says the paper comes from the editor, and a profile that mentions the paper is a leak. The app reads the same settings files; whether it adds more (claude.ai preferences, memory) is untested.
- The check is built in: the last line of every report answers “What do you know about who wrote or sent this?” Read a few out in the round-up.
- Separate sessions, not subagents: the room can see them, and a main session briefing subagents would write their prompts itself (“the user’s paper”). What a subagent sees wasn’t tested.
- On
ai-label-paper: the package step should catch every anonymisation slip (KEY item 9), including the PDF’s author field and the self-citation. Good referees find the missing interaction test, the manipulation check before the outcome, and the mediator measured after it. - The comparison step is the sycophancy talk made personal: expect praise first and no verdict (our measurement).
- Skill hook: “Turn what we just did into a skill called referee-panel”. The one skill worth making today.
What it can touch
About 20 minutes, while the three referees run. Talk first; the fence exercise is optional (handout section 7).
- The stance, plainly: lots of access, on purpose. The 50th “yes” is never read; git makes mistakes cheap; auto mode stops what matters. Restrict the few things that matter, mostly by keeping them out of the folder.
- In the app: the mode is a dropdown under the prompt box (Manual, Accept edits, Plan, Auto; there’s also Bypass permissions: don’t mention it). Which mode the Code tab starts in wasn’t checked.
- What we tried (20 October, terminal versions: Claude Code 2.1.302, Codex 0.162; the apps run the same agents):
- Claude, manual: reads in the folder; reading elsewhere, writing, deleting and web pages all ask.
- Claude, auto: edits, Python, reading outside the folder (asks once), a web page,
rm -rf data, evengit reset --hard && git clean -fdxwent ahead. Posting a file to an outside address was stopped: “denied by the Claude Code auto mode classifier. Reason: [Data Exfiltration]”. - Codex sandbox (Linux): reads anywhere, writes in the project and
/tmp, network off. - The deny rule
Read(./interviews/**)blocked Claude’s Read,cat,cpand a Python one-liner, butgrep -rprinted a transcript. Hence “a fence, not a wall”.
- Sources (official docs, checked 20 October 2026): permission modes and auto mode, code.claude.com/docs/en/permission-modes; settings and rule syntax, /permissions and /settings; “On native Windows, Claude Code runs commands unsandboxed”, /sandboxing; training covers Claude Code, retention 5 years on, 30 days off, /data-usage; Codex presets and the Windows sandbox, learn.chatgpt.com/docs/sandboxing; ChatGPT training controls apply to Codex, help.openai.com.
- Handle with care: “it sends what it reads” is the honest short version, not a guarantee. On personal data, “your consent forms and data plan decide” is safe; don’t claim what BI’s rules say unless you’ve checked.
Any time
The handout has all of these (sections 8–17), and each has a deck. Use them for fast people, for questions, or not at all.
Git: save points
The full deck (git, 25 slides, four exercises) was Strategy’s mandatory block. Today git is one prompt in setup and a habit in the cleanup and the paper; the deck is for anyone who asks “how is this different from OneDrive?”
- The answer, in one line: OneDrive is a backup, git is a lab notebook. Keep both, but not git inside OneDrive.
- What a save point is is the slide people missed in Finance: the files you see are always the current version; a commit copies them into
.git; a revert copies a save point back. - Sources (checked 20 October 2026): Torvalds wrote git in April 2005 (Pro Git, §1.2); Stack Overflow Developer Survey 2022, professional developers: Git 96.65%; Windows on git, about 300 GB (Brian Harry, Microsoft DevBlogs, May 2017); Claude Code checkpoints “do not track files modified by Bash commands” (code.claude.com, Checkpointing). That last one is why git, not the app’s own undo, protects the cleanup.
What it knows about you
- The profile matters today in one way: it’s read in every session, the referees’ too. If someone has a profile that names their current paper, they switch it off for the panel.
- The schools question for marketing (“grow by targeting a loyal niche, or by reaching as many category buyers as possible?”, Ehrenberg-Bass vs. brand equity) is untested. Strategy’s version changed the framing, not the verdict; expect the same.
- The Berger demo is the tested example of one line changing the school. Demo on Claude: Codex ignored one-line profiles in testing.
- The deck (
knows) still opens with Strategy’s “your field, as I imagine it”; it needs marketing cards before anyone sees it.
References
- Search: be specific. Deep searches are slow and expensive on usage; show, don’t run.
- DOI check: the practice
.bibhas six traps (key:scratch/examples/references/README.md). Onai-label-paper, the fake reference is Bergstrøm & Lind (2023): its DOI belongs to an unrelated 2024 paper. - Claim check: the centrepiece, on their own draft.
- At Strategy, people were lukewarm about this block (debrief, note 4), so it’s here, not in the afternoon.
Words you’ve heard
Nine tiles: skills, connectors, subagents, the project file, plan mode, context and /compact, worktrees, GitHub, agents in the cloud. The handout now says where each lives in the app (plan mode in the mode control, Worktree before a session, Cloud under the prompt box), from the desktop docs; untested.
Leave a trail
About 50 minutes; the key is scratch/examples/leave-a-trail/README.md.
- The practice folder (
props/messy-project.zip): three willingness-to-pay outliers deleted by hand in Excel, no record. With them, the effect flips sign and stops being significant (N = 223, −292 NOK, p = 0.26; without, N = 220, +53 NOK, p = 0.02). Only the rerun test finds it. - Reinhart and Rogoff, accurately: −0.1% reported above 90% debt to GDP; Herndon, Ash and Pollin got +2.2%. The Excel range error accounts for only about 0.3 points; the rest is excluded years and the weighting. Don’t say the Excel error alone flipped it.
- The test is now a new session, not
/clear; the same thing in the app.
Notes and structure
The four questions (from the tidy-project skill), the paper folder, dev/ vs. local/, the log habit. Lighter than the rest.
Your own work
- The power-check card (2 × 2, d = 0.3 vs. 0.1, power for the interaction) is the one for this room: it’s the
ai-label-paperproblem, in their own design. Untested; expect a few hundred per cell, which is the point. - Close: “What will you try on Monday?” Three answers out loud.
Interviews
Strategy’s qualitative block: eight fictional founder interviews (props/pivot-interviews.zip; key: scratch/examples/pivot-interviews/KEY.md). Three pseudonymisation slips, a contradiction, a planned-or-emergent story, a case that doesn’t fit. Tested 21 October with Codex: all found, every one of 187 quotes matched. For the qualitative people, if they ask.
GitHub, the project file, connectors
If asked. Decks github, agents, mcp; handout sections 13, 16 and 17. One project file, AGENTS.md, for both tools (Claude reads it when there’s no CLAUDE.md). GitHub setup is one prompt once they have an account.
Questions you’ll get
The longer prepared answers are in scratch/august-course/grievance-answers.md. The short versions:
- “Is it OK to upload my unpublished paper?” It’s your paper: training off, co-authors told, and data you can’t share left out. That’s the privacy slide in Your paper.
- “Can I use this when I referee someone else’s paper?” A different question. Many publishers ask referees not to upload manuscripts they review into AI tools. Check the journal’s policy. Your own paper is yours.
- “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. 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 paper is instructive: they find different problems.
- “Will the department pay?” Yes: Marketing reimburses a paid Claude or ChatGPT subscription. Receipt to Christine.
- “Can I get the slides?” Yes: everything is on vibing.jonasmoss.com → Marketing, and the handout has every prompt.