Generative AI for Scholarship
Harvard Data Science Initiative (HDSI) & Faculty of Arts and Sciences (FAS)
In this exercise you will use Claude Code to analyze real thermal data from the Vera C. Rubin Observatory primary mirror. The dataset contains temperature measurements at 15-minute intervals from June through December 2025, recorded at the summit in Chile.
First, create a subdirectory for this session's work and navigate to it:
cd ~/GAIsandbox
mkdir -p session3
cd session3Now download the data file:
curl -O https://astrostubbs.github.io/GenAI-for-Scholarship/data/session3/rubin_mirror_temps.csvThe file contains about 17,600 rows with two columns:
timestamp Date and time of the measurement (UTC), at 15-minute intervalstemperature Mirror temperature (°C)
Start Claude Code in your ~/GAIsandbox/session3 directory:
claudeWork through the following steps. For each one, type the prompt (or your own version of it) into Claude Code and observe what it does.
Try a prompt like:
"Look at the file rubin_mirror_temps.csv. Tell me the column names
and show me the first few lines of data. Are there any anomalies
in the data file?"Notice how Claude Code reads the file, summarizes its structure, and flags any issues (missing values, gaps in the time series, etc.) without you writing any code.
"Write a Python program to plot temperature vs. time from
rubin_mirror_temps.csv. Label the axes with units and give
the plot a descriptive title."Claude Code will write a Python script, run it, and display the resulting plot. Look at the code it wrote — is it what you would have written?
"Add a histogram of the temperature values. Show it as a
separate figure."This builds on the previous step. Notice how Claude Code remembers the context of what it has already done.
"When did the maximum and minimum temperatures occur?
Report the dates, times, and values.""Modify the plotting program to show only the most recent
week of data. For each day, estimate the time of sunset at
Cerro Pachon, Chile (latitude -30.24, longitude -70.74) and
mark it on the time series plot with a large red dot."
This is a more complex request — Claude Code needs to figure out how
to compute sunset times (it will likely use an astronomy library like
astropy or ephem), interpolate the
temperature at that time, and add the markers to the plot. Watch how
it breaks the problem into pieces.
"Take the Fourier transform of the temperature time series
and report the dominant periods, in days. Plot the power
spectrum."Look at the periods Claude Code reports as dominant. Do they make physical sense? What period would you expect to see most prominently for outdoor temperature data? Is the answer you got consistent with that expectation?
Pause here. Look at the dominant periods Claude Code reported. Are we asking the right question of the data? Discuss with your neighbors — what would you expect to see, and does the answer match? If not, think about why, and then ask Claude Code a better question.
Claude Code automatically saves your conversation history in the
~/.claude/ directory, organized by the working directory
where you ran it. Let's test this.
First, quit Claude Code by typing /exit or pressing
Ctrl-D. Then restart it:
claude -r
You'll see a list of your recent sessions. Select the one for
~/GAIsandbox/session3. The session resumes with full
context — you can continue asking questions about the data, request
modifications to the plots, or reference anything from earlier in the
conversation. Try asking:
"What was the maximum temperature we found earlier?"Claude Code remembers. This makes it easy to work in stages — do some analysis, take a break, come back later and continue where you left off.
Good research practice means documenting what you did. Ask Claude Code to write up the session as a self-contained HTML lab notebook:
"Create a NEW HTML file called lab_notebook.html that documents
the analysis we just did. Include:
- A title with today's date
- A summary of each analysis step and what we found
- All the plots, embedded directly in the HTML as base64 images
- Links to each Python source file used
- A section at the end listing the data file and its provenance
The HTML should be self-contained — openable in any browser
with no external dependencies."
Open lab_notebook.html in your browser to see the
result. This is a reproducible record of your analysis session.
For complex, multi-step tasks, Claude Code offers Plan Mode. In Plan Mode, Claude explores your codebase and designs an implementation approach first, presenting you with a detailed plan before executing anything. This is especially useful for tasks that require:
/plan).
The mode indicator in the status area will change to show you're in Plan Mode.Think of Plan Mode as the "design" phase and normal mode as the "implementation" phase. This two-phase workflow prevents Claude from diving into implementation before you've agreed on the approach.
Now let's use Plan Mode for a challenging task: predict the temperature at sunset, three hours ahead of time, by comparing multiple machine learning methods.
Here's the prompt to use in Plan Mode:
I want you to act as an expert in machine learning methods. The goal is
to predict the temperature at sunset on any given day, using the entire
time history of temperature measurements up to 3 hours before sunset.
For each day:
- Input: all temperature measurements from the start of that day up to
3 hours before sunset
- Target: the actual temperature at sunset
Break the data file into 2 sections: use even days of the month for
training, and odd days of the month for testing and validation.
The metric to optimize is MAE (mean absolute error) between predicted
and actual sunset temperature.
First, suggest several approaches that might work well for this problem,
considering that we have rich time series history as input. Then
implement and compare at least 3 different methods. Make a quantitative
performance comparison and give me a recommendation on what I should use.In Plan Mode, Claude will explore the data, think through the approach, and present you with a detailed plan covering steps like:
Review the plan. If it looks good, approve it and exit Plan Mode (press Shift+Tab again). Claude will then execute the plan: installing packages, writing code, training models, and producing comparison plots and metrics.
This demonstrates the full power of an agentic coding assistant on a real research problem — planning, implementing, and evaluating a solution end-to-end.
After running the ML comparison, you'll have two PNG files with results:
ml_comparison_v2.png (model performance comparison) and
ml_residuals_v2.png (comprehensive residual analysis).
Now create a professional HTML report to document your findings.
Ask Claude Code:
"Create a NEW HTML file called ml_summary.html that presents the machine
learning comparison results. This should be a separate report from the
lab notebook. Include:
1. Title: 'Machine Learning for Sunset Temperature Prediction'
2. Introduction: Brief description of the prediction problem and why it matters
3. Methods: Summarize each of the 5 ML approaches tested and what they do
4. Results: Embed both PNG files (ml_comparison_v2.png and ml_residuals_v2.png)
5. Performance table: Show MAE, RMSE, and R² for each method
6. Recommendation: Which method should be used and why?
7. Interpretation: What do the residual plots tell us about model quality?
The HTML should be self-contained with embedded images and professional
styling."This creates a polished summary suitable for sharing with collaborators or including in research documentation. Notice how Claude Code can take the raw analysis results and transform them into a publication-ready format.
Claude Code saves all sessions locally in ~/.claude/,
organized by the directory where you ran it. This means:
claude -c — continues the most recent session in the current directoryclaude -r — shows a picker to resume any previous session
The session files themselves are in ~/.claude/sessions/
as JSON files, but the intended way to access them is through
claude -c and claude -r.
Since Claude Code sessions are tied to the working directory, use your directory structure to organize different projects and analyses. For example:
~/GAIsandbox/
session1/ # Session 1 work
session2/ # Session 2 work
session3/ # This session - thermal analysis
my-project/ # Your own research project
When you navigate to a project directory and run claude -c,
you automatically resume in the context of that project. This keeps
different analyses separate and makes it easy to switch between them.
Think of each directory as a separate "lab notebook" with its own
conversation history. For this exercise, all work will be in
~/GAIsandbox/session3/, keeping it separate from other sessions.
Claude Code has several built-in commands (slash commands) that are useful during your session. Type them at the prompt:
/help Show help and available commands/permissions View and manage tool permissions (Bash, Edit, Write, etc.)/cost Show API cost for the current session/model Show which model you're using (opus, sonnet, haiku)/plan Enter Plan Mode (same as Shift+Tab)/exit Exit Claude Code (same as Ctrl-D)/clear Clear the conversation history/tasks Show task list if you're using task tracking
The /permissions command is particularly useful if Claude Code
gets blocked on certain operations. You can grant permissions for specific
tools or prompts during your session.
After completing the exercises, it's helpful to see how much API usage they consumed. In Claude Code, type:
/costThis will show you the total API cost for your current session, broken down by model usage and token counts. This helps you understand the typical costs of using Claude Code for research tasks and plan your budget accordingly.
Reference implementations, teaching notes, and solutions are available on a separate page:
Please take a moment to share your feedback — it helps us improve future sessions.