Week 06: From Data to Report

Vibe research vs building an economics-quality report with CLI tools

Published

February 10, 2026

Vibe research vs building an economics-quality report with CLI tools

Overview

This week ties together everything from weeks 4-5: you’ll use Claude Code to download real data, explore it, and produce a polished PDF report – all from the terminal. The twist: we start by letting AI loose (“vibe report”) and then compare that to a carefully directed, economics-quality report. The contrast is the lesson.

We use the US Earnings (CPS MORG) case study – real Current Population Survey data on wages, education, occupation, and demographics. Read more on CPS here.

Learning Outcomes

By the end of the session, students will:

Download and prepare a real dataset from an online repository using CLI tools
Experience the difference between undirected (“vibe”) and directed AI output
Iteratively refine graphs from basic to publication-quality
Produce a constrained, economics-style PDF report with exhibits and regressions

Preparation / Before Class

Prerequisites

Required:

Working Claude Code installation from Week 04
Python environment with pandas, matplotlib, seaborn, statsmodels installed
Familiarity with running Claude Code in a project folder (Week 05)

Review:

US Earnings case study – browse the data description and key variables
Installing CLI Tools – if you need to catch up

Useful background:

The earnings data comes from the case study on gender and age differences in earnings

Class Material

Part 1: Get the Data (15 min)

The goal: Use Claude Code to download the full CPS MORG dataset from OSF, understand its structure, and prepare it for analysis. This showcases real CLI data work – no browser, no manual downloads.

Step 1: Download from OSF

Open Claude Code in your project folder and try:

Download the CPS MORG 2014 earnings data from OSF.
The dataset page is at https://osf.io/g8p9j/files/4ay9x
Also get the codebook PDF from https://osf.io/uqe8z as well as occupation codes from https://osf.io/g8p9j/files/57n9q
Save all to a data/ folder.

Claude Code will use curl or wget to fetch the files directly. Watch how it handles the download, checks file sizes, and confirms the data arrived.

Step 2: First look

Read in the CSV. How many rows and columns?
Show me the first few rows and basic summary stats.

The full MORG file is large. Let’s focus on one.

Create a frequency table of states. Drop small ones. Make it pretty.

Pick a state and add to script:

Filter to the state [your choice]. Print the state code. 
Save as morg-2014-emp-state5.csv.

Check the state code in pdf.

Step 3: Variable dictionary

This one (bit sleazy) promp, can try something better

This is my data and the codebook. Create a variable dictionary. Use the pdf i shared earlier. Output as markdown. For each variable: varname, labels, type, coverage (% missing), mean and mode. Round up numbers. Look at cps and provide short labels. Get me an .md I can download.

Could you refine the prompt to make output better, more structured, or simpler?

Part 2: The Vibe Report (20 min)

The experiment: Give AI a vague, open-ended prompt and see what happens.

Create a nice looking report on an interesting question
using the content of the data folder only.

Let Claude Code run. It will likely produce a long, generic report with many graphs, broad research questions, and decorative formatting.

Code to pdf

Make sure a pdf is created. It may install bits – here is a guide to create pdf docs directly with AI.

Class discussion: Evaluate the vibe report?

Look at what was produced and score (1/10, where 10 is best in class, paper or blogpost) and discuss:

Length – does everything useful and relevant?
Is the research question well defined?
Graphs and tables: good, informative, well labeled?
Economics rigor: did it create exhibits that are informative and useful? Does it use the appropriate tools (descriptive table, regression, etc)
Descriptions: Is the text describes the exhibits and results well? Does it interpret them correctly? Does it make the right conclusions?

Overall, what do you think about the choices it made?

code

Check the code. How did it mix analytics and text?

Part 3: Some focused Exercises (30 min)

Now let’s do it properly. Go back and build up the analysis step by step, iterating on some exhibits.

Task 1: Heatmap

Show a heatmap of hourly wages by occupation and education level. Use viridis.

Iterate to make it good (if needed)

Task 2: Adding analysis: Now add regression analysis to support the graphs.

Run an OLS regression of hourly wages on gender, controlling for
age, education (grade92), and occupation (occ2012).
Report the results in a clean table.

Iterate on the specification:

Add age squared. Report robust standard errors.
Interpret the gender coefficient -- what does it mean in dollar terms?

Prompting Tips for Data Analysis

What We Learned About Prompting

This week demonstrates several prompting principles in action:

Be specific about your tools and preferences

“Using Python with pandas and matplotlib…”
“Use viridis color scheme” – state your preferences explicitly
“Save as PDF” – specify output format

Include data structure information

Upload the data or describe it: “DataFrame with columns: age, sex, earnwke, grade92, occ2012”
Point to the codebook: “Use the PDF codebook I shared”
Specify the unit: “hourly wages, computed as earnwke / uhours”

Think constraints on output

“3-5 exhibits, no more”
“600-800 words”
“Professional formatting with numbered exhibits”

Vague vs. specific prompts – a comparison

Vague: "Analyze this earnings data and write a report"

Specific: "Using CPS MORG data for California, estimate the gender wage gap controlling for education and occupation. Create a LOESS plot of residual wages by age and gender. Write a 700-word report with 4 exhibits."

The vague prompt gives you something. The specific prompt gives you what you need. You’ll test it in the assignment.

Discussion Points

When is a “vibe report” actually useful? (Exploration, brainstorming, first look at data)
How do you decide how many exhibits belong in a report?
What’s the right balance between letting AI explore freely vs. directing every step?
How would this workflow differ if you were using R instead of Python?

Assignment

Assignment 6: From Data to Report

Due: Before Week 7

Use Claude Code to download a dataset, explore it, and create a focused PDF report.

Full Assignment Details

Resources

Case study: US Earnings (CPS MORG) – data files and documentation

Graph walkthrough: Creating Graphs – one example with R.

Data source: OSF - CPS MORG data | Codebook PDF

Textbook reference: Chapter 9A - Gender and age differences in earnings

--- title: "Week 06: From Data to Report" subtitle: "Vibe research vs building an economics-quality report with CLI tools" date: "2026-02-10" --- ::: {.hero-section} ::: {.container} ::: {.hero-title} Week 06: From Data to Report ::: ::: {.hero-subtitle} Vibe research vs building an economics-quality report with CLI tools ::: ::: ::: --- ## Overview This week ties together everything from weeks 4-5: you'll use Claude Code to **download real data**, **explore it**, and **produce a polished PDF report** -- all from the terminal. The twist: we start by letting AI loose ("vibe report") and then compare that to a carefully directed, economics-quality report. The contrast is the lesson. We use the **US Earnings (CPS MORG)** case study -- real Current Population Survey data on wages, education, occupation, and demographics. Read [more on CPS here](https://gabors-data-analysis.com/datasets/#cps-earnings). ### Learning Outcomes By the end of the session, students will: - Download and prepare a real dataset from an online repository using CLI tools - Experience the difference between undirected ("vibe") and directed AI output - Iteratively refine graphs from basic to publication-quality - Produce a constrained, economics-style PDF report with exhibits and regressions ## Preparation / Before Class ::::: {.week-card .card} ::: card-header **Prerequisites** ::: ::: card-body **Required:** - Working Claude Code installation from [Week 04](../week04/index.html) - Python environment with `pandas`, `matplotlib`, `seaborn`, `statsmodels` installed - Familiarity with running Claude Code in a project folder (Week 05) **Review:** - [US Earnings case study](/case-studies/earnings/) -- browse the data description and key variables - [Installing CLI Tools](../da-knowledge/install-cli.html) -- if you need to catch up **Useful background:** - The earnings data comes from the case study on [gender and age differences in earnings](https://gabors-data-analysis.com/casestudies/#ch09a-estimating-gender-and-age-differences-in-earnings) ::: ::::: ## Class Material ::::: {.week-card .card} ::: card-header **Part 1: Get the Data (15 min)** ::: ::: card-body **The goal:** Use Claude Code to download the full CPS MORG dataset from OSF, understand its structure, and prepare it for analysis. This showcases real CLI data work -- no browser, no manual downloads. **Step 1: Download from OSF** Open Claude Code in your project folder and try: ``` Download the CPS MORG 2014 earnings data from OSF. The dataset page is at https://osf.io/g8p9j/files/4ay9x Also get the codebook PDF from https://osf.io/uqe8z as well as occupation codes from https://osf.io/g8p9j/files/57n9q Save all to a data/ folder. ``` Claude Code will use `curl` or `wget` to fetch the files directly. Watch how it handles the download, checks file sizes, and confirms the data arrived. **Step 2: First look** ``` Read in the CSV. How many rows and columns? Show me the first few rows and basic summary stats. ``` The full MORG file is large. Let's focus on one. ``` Create a frequency table of states. Drop small ones. Make it pretty. ``` Pick a state and add to script: ``` Filter to the state [your choice]. Print the state code. Save as morg-2014-emp-state5.csv. ``` Check the state code in pdf. **Step 3: Variable dictionary** This one (bit sleazy) promp, can try something better ``` This is my data and the codebook. Create a variable dictionary. Use the pdf i shared earlier. Output as markdown. For each variable: varname, labels, type, coverage (% missing), mean and mode. Round up numbers. Look at cps and provide short labels. Get me an .md I can download. ``` Could you refine the prompt to make output better, more structured, or simpler? ::: ::::: ::::: {.week-card .card} ::: card-header **Part 2: The Vibe Report (20 min)** ::: ::: card-body **The experiment:** Give AI a vague, open-ended prompt and see what happens. ``` Create a nice looking report on an interesting question using the content of the data folder only. ``` Let Claude Code run. It will likely produce a long, generic report with many graphs, broad research questions, and decorative formatting. **Code to pdf** Make sure a pdf is created. It may install bits -- here is a guide to [create pdf docs directly](/da-knowledge/pdf-guide.html) with AI. **Class discussion: Evaluate the vibe report?** Look at what was produced and score (1/10, where 10 is best in class, paper or blogpost) and discuss: - Length -- does everything useful and relevant? - Is the research question well defined? - Graphs and tables: good, informative, well labeled? - Economics rigor: did it create exhibits that are informative and useful? Does it use the appropriate tools (descriptive table, regression, etc) - Descriptions: Is the text describes the exhibits and results well? Does it interpret them correctly? Does it make the right conclusions? Overall, what do you think about the choices it made? **code** Check the code. How did it mix analytics and text? ::: ::::: ::::: {.week-card .card} ::: card-header **Part 3: Some focused Exercises (30 min)** ::: ::: card-body **Now let's do it properly.** Go back and build up the analysis step by step, iterating on some exhibits. **Task 1: Heatmap** ``` Show a heatmap of hourly wages by occupation and education level. Use viridis. ``` Iterate to make it good (if needed) **Task 2: Adding analysis:** Now add regression analysis to support the graphs. ``` Run an OLS regression of hourly wages on gender, controlling for age, education (grade92), and occupation (occ2012). Report the results in a clean table. ``` Iterate on the specification: ``` Add age squared. Report robust standard errors. Interpret the gender coefficient -- what does it mean in dollar terms? ``` ::: ::::: ## Prompting Tips for Data Analysis ::::: {.week-card .card} ::: card-header **What We Learned About Prompting** ::: ::: card-body This week demonstrates several prompting principles in action: **Be specific about your tools and preferences** - "Using Python with pandas and matplotlib..." - "Use viridis color scheme" -- state your preferences explicitly - "Save as PDF" -- specify output format **Include data structure information** - Upload the data or describe it: "DataFrame with columns: age, sex, earnwke, grade92, occ2012" - Point to the codebook: "Use the PDF codebook I shared" - Specify the unit: "hourly wages, computed as earnwke / uhours" **Think constraints on output** - "3-5 exhibits, no more" - "600-800 words" - "Professional formatting with numbered exhibits" **Vague vs. specific prompts -- a comparison** *Vague:* `"Analyze this earnings data and write a report"` *Specific:* `"Using CPS MORG data for California, estimate the gender wage gap controlling for education and occupation. Create a LOESS plot of residual wages by age and gender. Write a 700-word report with 4 exhibits."` The vague prompt gives you *something*. The specific prompt gives you what you *need*. You'll test it in the assignment. ::: ::::: ## Discussion Points - When is a "vibe report" actually useful? (Exploration, brainstorming, first look at data) - How do you decide how many exhibits belong in a report? - What's the right balance between letting AI explore freely vs. directing every step? - How would this workflow differ if you were using R instead of Python? ## Assignment ::: {.callout-note icon="false"} ## Assignment 6: From Data to Report **Due:** Before Week 7 Use Claude Code to download a dataset, explore it, and create a focused PDF report. [Full Assignment Details](../assignments/assignment_06.html){.assignment-badge} ::: ## Resources **Case study:** [US Earnings (CPS MORG)](/case-studies/earnings/) -- data files and documentation **Graph walkthrough:** [Creating Graphs](/week06/assets/creating-graphs.html) -- one example with R. **Data source:** [OSF - CPS MORG data](https://osf.io/4ay9x) | [Codebook PDF](https://osf.io/uqe8z) **Textbook reference:** [Chapter 9A - Gender and age differences in earnings](https://gabors-data-analysis.com/casestudies/#ch09a-estimating-gender-and-age-differences-in-earnings)