Two days, three levels of letting go.

Level 1 · Chatbot · we are the loop

1. 01

   Where we actually are

   How capable the tools are, and what happens when we ask.

   How capable these tools have become, and how quickly it happened. What the benchmarks measure, and what they do not. We also cover what actually happens when we ask a question about a dataset: the model does not compute statistics itself; it writes code, and a sandbox runs that code.

   Four years in one timeline, reading benchmarks, a first analysis in a minute, three levels of delegation.

2. 02

   Models, modes and apps

   Which model, mode and app, and when slower is worth the wait.

   Which model to open and when the slower one is worth the wait, what the different modes are for, and how the browser and desktop versions differ.

   Which model?

   Top tier for real work: GPT-5.6 sol, Claude Opus. Lighter tiers for chores: luna, tera, Haiku. Harder question → let it think longer.

   Which mode?

   Regular for back-and-forth, extended thinking for hard problems, deep research for long sourced reports. Code, files and search are tools it picks up on its own.

   Which app?

   Web (claude.ai, chatgpt.com) → Level 1. Desktop and terminal (Claude Code, Cowork, Codex) → Levels 2 and 3. Same brains, different reach.

   Fast vs thinking models, extended thinking, deep research, code execution, web vs desktop app, Claude and ChatGPT side by side.

3. 03

   Prompting

   A four-part checklist, then one question asked badly and well.

   A short checklist (context, goal, output, done) covers most cases. We then ask the same question badly and well on the same data, and compare what comes back.

   The four-part checklist, weak vs specified, asking for the code, iterating on an answer.

   

   cold startafter two roundswith a skill

   “Can you create a scatter plot that shows the relationship between GDP per capita and happiness in 2023.” The default answer.The same prompt plus seven requests: X markers colored by happiness, a dashed OLS line, a gray 95% band, magma, the legend top right.The same prompt plus one line: “Use my watercolor-graphs skill.”

4. 04

   Setting up a project

   A folder, its data, and instructions that carry over.

   Where the work lives: a folder, the data inside it, and standing instructions that carry across sessions. That way we are not re-explaining our setup every time.

   Create a project, add reference files, standing instructions, reuse across sessions.

5. 05

   Getting data in

   FRED and World Bank by API, files of every format, merged.

   Pulling series directly from an API rather than looking for a download button. Loading files in the formats data usually arrives in, and combining sources into one dataset.

   FRED API, World Bank API, CSV and Excel, PDF tables, merging data.

6. 06

   Cleaning and reshaping

   Long to wide, missing values, outliers, each in a sentence.

   The unglamorous half of every project. Each step is described in a sentence rather than written by hand.

   Long ↔ wide, missing values, outliers, duplicates, recode and label, build new variables, filter samples.

7. 07

   Describing the data

   Tables for a paper, and what an average is hiding.

   Summary tables of the kind we would put in a paper. Plus the breakdowns that show whether an average is hiding something.

   Summary tables, by group, correlation matrix, cross-tabs, balance tables, export to Word.

8. 08

   Figures and maps

   From a raw file to a figure ready to publish.

   Going from a raw file to a figure that is ready to publish. Then what each type of chart shows well, and what it leaves out.

   Trends over time, rankings, distributions, scatter with fit, world map.

9. 09

   Regression

   OLS, controls, fixed and random effects, on one panel.

   Estimation on the same panel, starting simple. Then we add the structure that makes the question interesting.

   OLS with confidence bands, controls, fixed effects, random effects, interpreting the output, regression tables.

   β = 0.713 per log point, about 0.49 per doubling · R² = 0.60 · n = 162. From the course data. The dot at far left is Venezuela: a break in its GDP data, not a poor, happy country.

10. 10

    Beyond the regression table

    Mediation, SEM, clustering, and any package we install.

    Methods that go past a single equation, including installing a library the sandbox does not come with. That is the point where most people realize any method with a Python package is available.

    Mediation, SEM, clustering, install a package.

11. 11

    Going further if we have time

    An event study around COVID, and wellbeing types.

    An event study around COVID, with the parallel-trends check. And a latent profile analysis that sorts countries into wellbeing types instead of ranking them on one line.

    Event study, parallel trends, latent profile analysis.

12. 12

    A first dashboard

    The analysis, as something others can open and click.

    Turning the analysis into something interactive that someone else can open and click through.

    Interactive dashboard, filters, export and share.

    

    A real deliverable, from the course's own folders.

Level 2 · Coworker · we are at lunch

13. 13

    Folder, goal, permission

    Point Claude Code, Codex or Cowork at a folder, with a goal.

    Claude Code, Codex and Cowork all work the same way: we point them at a project folder, give them a goal, and grant permission. They then plan the work, split it across parallel agents, and return a finished deliverable with the log attached.

    A tour of the desktop app, open a folder, permissions, watch it plan, read the log.

    A look ahead, in ChatGPT: Codex. The menu at the top left switches the desktop app to Codex. Sites, Scheduled and Plugins are the same ideas as Claude's, some under different names.

    

    

14. 14

    The toolkit

    CLAUDE.md, memory, skills, sub-agents, connectors, plugins.

    The vocabulary that makes a coworker different from a chat window, each part demonstrated. CLAUDE.md house rules, persistent project memory, reusable skills, sub-agents, MCP connectors, and plugins.

    CLAUDE.md, project memory, skills, sub-agents, MCP and connectors, plugins.

15. 15

    What we build

    Reports, websites and dashboards, and one honest failure.

    Real projects, start to finish, each one handed over whole rather than prompted a step at a time. And one that fails, because working out why teaches more than any success does.

    The full report, one more sentence: a website, the Stata audit from the terminal, the multiverse, text as data, the economic time machine on live FRED data, the Carbon Divide (a World Bank story), and “write my paper in my voice”: a failure.

    the full report a website the multiverse a time machine a data story “write my paper in my voice”a failure

    Each one opens full size.

Level 3 · Manager · we design the loop

16. 16

    Designing the loop

    The coworker, promoted: we set the goal, the check, the stop.

    Not a different machine: the same coworker, promoted to manager. We design the loop: what done means, how to check it, and when to stop. The manager splits the goal into jobs and sends out an army of agents; every one of them works in a loop, and nothing ships until it passes the check. That is more control, not less. /goal keeps the work going until our condition is met; schedules and triggers can start a run, but only the check finishes the work.

    An army of agents, /goal, goal · check · stop, checking in from anywhere.

Our part · checks and judgment

17. 17

    Verification design

    Nothing crashes, the chart looks fine, the answer is wrong.

    The worst errors are silent: nothing crashes, the chart looks fine, and the answer is still wrong. The models are now good checkers of their own work and of each other's, so the skill moves from spotting every error by hand to designing the checks. We decide what must be verified, write the check down, and hand it to a reviewer agent: a rerun from raw data, a second model cross-examining the first, a reconciliation against a number we trust. We weigh how easily something can be checked against how costly a mistake would be, then work through a real one: a plausible claim that the world got 0.83 points sadder, three files, and a silent error (the survey grew from 27 countries to 138, and the same 27 fell just 0.29).

    Silent errors, what must be checked, writing the check down, reviewer agents, a second model cross-examining, reconciling against a known number, finding the silent error.

    the claimthe check

    

    Change in average life satisfaction, 2005 → 2023

    every country surveyed27 → 138 countries−0.83

    the same 27 countries27 → 27 countries−0.29

    0−0.3−0.6−0.9

    The claim: the world got 0.83 points sadder. The code ran, and the chart looks fine.The check: the same 27 countries fell just 0.29. About two-thirds of the “decline” is the survey growing, from 27 countries to 138.

18. 18

    Judgment

    One chart, four headlines: which would we sign?

    Beautiful is now the default: models design figures, dashboards and whole websites better than most of us. What is left is judgment: deciding what a result should say, to whom, and what it cannot claim. One chart, four headlines: which would we sign? Then we write our standards down once, as a skill, so that every session after applies them. The look of this website comes from one.

    One chart and four headlines, what a chart can claim, association vs cause, encoding our standards as a skill.

    Money buys happiness.don't sign Money explains 60% of happiness.don't sign Doubling a country's GDP adds half a point of happiness.don't sign Richer countries tend to rate their lives higher.sign it

    Pick a headline to see the verdict.

    This website was built with Claude, and its look comes from a skill like that.

19. 19

    The closing hour

    Is AI making us stupid? What's left for experts?

    Where this leaves us. Is AI making us stupid? Are we going to lose our jobs? What's left for experts? What changes about how we work, what does not, and what we would refuse to delegate. Bring an opinion.