This guide is built to be handed to an AI assistant, which then walks you through it one step at a time. Claude Code in a terminal, Claude Desktop, and Codex all work. Save this page to your machine, open a fresh assistant session, give it this file, and paste this:
Walk me through Session 2.4 one step at a time.
Give me only the current step. Stop after each step and wait for me to say it's done.
Before each step, check whether it's already done on this machine, and skip it if it is.
Tell me what each command does before I run it.
Start at the beginning.Then work through it together. Tell the assistant when a step is finished, or paste the error when one isn’t.
Assistant: the rules below govern this whole session. They override any default habit of reading a document and summarizing it.
CLAUDE.md or
AGENTS.md records which one; if it does not, ask once and
do not ask again. Where a step below has a ### Claude Code
version and a ### Codex version, deliver only theirs and do
not mention the other exists. Everywhere else the work is identical, so
say “your assistant” rather than assuming Claude. This is about the tool
the learner installs and works in, not about which assistant is reading
this guide.About 60 minutes. Run this from ~/ai-training, the
working directory you made in Session 2.1.
Posture: public, synthetic, or personal data only. Today’s data is FRED (Federal Reserve Economic Data) — public, downloadable, well-documented. If you’d rather use a different public dataset (Our World in Data, BLS, World Bank, a public survey), substitute freely; the workflow is the same.
By the end of this session you will have, inside
~/ai-training/exhibits/<topic-slug>/:
data/raw.csv — a public CSV pulled from FRED, with a
known SHA-256 hash recorded in data/manifest.txt.make_plot.py — a small matplotlib script Claude wrote
with you that turns the CSV into a figure.figure-v1.png — the first cut.figure-v2.png — second cut, with NBER recession bars
added.figure-v3.png — third cut, with your style applied
(palette, font, aspect ratio).~/ai-training/MEMORY.md capturing
your figure preferences.Three figures, one script, one tagged dataset, three remembered preferences. The figures are the visible output; the memory entries and the script are what make the next plot 5 minutes instead of 45.
A production version of this is a figures session that pulled survey response data from an Excel file and produced exhibits constrained by preferences from a previous session: 4:3 aspect ratio so two figures fit on a slide, Times New Roman 12pt black labels, no gridlines, single-select questions plotted as pie charts and multi-select as horizontal bar charts. The preferences had been written to MEMORY.md after a prior session, so the figures came out right on the first try — no re-explaining the house style. By the end of this session you’ll have written down enough of your own preferences that the next figure starts where this one ended.
The trap in figure-making: you produce one chart that’s almost-right, eyeball it, ship it. Three chart sessions later you’ve spent four hours on chart aesthetics across three documents and no two figures look the same.
The unlock is the same as Session 2.3’s memo iteration, applied to images:
figure-v1.png,
figure-v2.png, figure-v3.png. The intermediate
cuts are cheap — saving them is free; recovering one when v3 went wrong
is invaluable.make_plot.py takes
the CSV and produces the figure in one command. Re-running with new data
is one line.The figure is the artifact. The script and the memory entries are what make the next figure cheap.
Already done? Run
python3 -c "import pandas, matplotlib; print('ok')". If it
prints ok, skip to Step 2.
If matplotlib and pandas aren’t installed:
pip3 install --user pandas matplotlibWhat this does. pip3 installs Python
packages, bundles of code someone else wrote that your script can
import instead of you writing it from scratch.
pandas reads and manipulates tables;
matplotlib draws figures from them. The --user
flag installs into your own home directory rather than the system Python
macOS ships with, so you never need sudo and can’t break a
tool the operating system depends on. Re-running this command is safe;
it just confirms the packages are already there.
Verify:
python3 -c "import pandas, matplotlib; print(pandas.__version__, matplotlib.__version__)"What this does. python3 -c runs the
string that follows as a one-line Python script, without saving anything
to a file, which is handy for a quick check like this one.
import pandas, matplotlib fails loudly if either package is
missing. print(...) reports which versions you have, useful
later if a script behaves differently on someone else’s machine.
Worth knowing. If pip3 install --user
puts binaries somewhere not on PATH, that’s fine for libraries.
python3 -c "import pandas" is the only check that matters
here.
Use the prompt from the top of this guide to start (or resume) the session. Tell Claude that matplotlib and pandas are already installed, so it can skip straight to the Reference Sections.
Claude takes over.
Claude: walk the user through pulling a CSV, verifying it, plotting it, and iterating. Peer-tone. Two rules:
Design: this session installs three habits — verify-then-plot, script-as-deliverable, preferences-to-MEMORY. Sessions 2.6 (corpus scoring) and 2.9 (PowerPoint deck) reuse the figures produced here. The 4:3 default is set today specifically because 2.9 pairs two figures per slide.
Before any data work:
python3 -c "import pandas, matplotlib" runs
cleanly.~/ai-training/MEMORY.md exists. If it doesn’t,
create it now with one section heading
## Figure preferences and an empty body — you’ll write into
it during Step E.Already done? Run
ls exhibits/unemployment-rate/data/raw.csv. If it exists,
skip to Step B.
The default dataset: U.S. unemployment rate, last 10 years, monthly,
from FRED. Series ID UNRATE. Public, no auth, downloadable
as CSV directly.
In Claude:
Create the directory exhibits/unemployment-rate/data/. Download the FRED
series UNRATE as CSV from
https://fred.stlouisfed.org/graph/fredgraph.csv?id=UNRATE
to exhibits/unemployment-rate/data/raw.csv. After downloading, record:
- SHA-256 hash of the file
- row count
- the first and last date in the file
- the column names
to exhibits/unemployment-rate/data/manifest.txt.Claude: use a shell tool to curl the URL, then a small Python snippet to compute the hash and read the date range. Don’t trust the file blindly — read the first 5 lines so the user can see the format. Confirm with the user that the data looks right (date column, value column) before moving on.
If the user wants a different dataset (BLS unemployment by state, OWID life-expectancy, World Bank GDP, a public survey CSV), substitute. The pattern is identical.
make_plot.py (10 minutes)Already done? Run
ls exhibits/unemployment-rate/figure-v1.png. If it exists,
skip to Step C.
Build the script with the user, in
exhibits/unemployment-rate/make_plot.py:
import pandas as pd
import matplotlib.pyplot as plt
from pathlib import Path
HERE = Path(__file__).parent
def load():
df = pd.read_csv(HERE / "data" / "raw.csv", parse_dates=[0])
df.columns = ["date", "rate"]
return df
def plot_v1(df, out):
fig, ax = plt.subplots(figsize=(8, 6))
ax.plot(df["date"], df["rate"])
ax.set_xlabel("Date")
ax.set_ylabel("Unemployment rate (%)")
ax.set_title("U.S. Unemployment Rate")
fig.tight_layout()
fig.savefig(out, dpi=150)
if __name__ == "__main__":
df = load()
plot_v1(df, HERE / "figure-v1.png")Run:
python3 exhibits/unemployment-rate/make_plot.py
open exhibits/unemployment-rate/figure-v1.pngWhat this does. python3 runs the script
end to end and writes figure-v1.png; open
launches it in Preview so you can see the result.
Read the figure together. It’s serviceable but plain. Common things to notice: aspect ratio is wrong for slides, no recession context, default fonts and colors. Make a list of what to fix.
Already done? Run
ls exhibits/unemployment-rate/figure-v2.png. If it exists,
skip to Step D.
NBER maintains the official U.S. recession dating. Public dates list at https://www.nber.org/research/business-cycle-dating. For the last 10 years, the relevant recession is February 2020 → April 2020 (COVID).
Add a plot_v2 function:
RECESSIONS = [
("2020-02-01", "2020-04-01"),
]
def plot_v2(df, out):
fig, ax = plt.subplots(figsize=(8, 6))
ax.plot(df["date"], df["rate"], color="#1f4e79", linewidth=1.5)
for start, end in RECESSIONS:
ax.axvspan(pd.Timestamp(start), pd.Timestamp(end), color="#cccccc", alpha=0.5)
ax.set_xlabel("Date")
ax.set_ylabel("Unemployment rate (%)")
ax.set_title("U.S. Unemployment Rate, with NBER recession shading")
fig.tight_layout()
fig.savefig(out, dpi=150)Update the __main__:
plot_v2(df, HERE / "figure-v2.png")Re-run, open figure-v2.png. The grey band is the recession.
Tell the user: “The recession bar is one tiny annotation that turns ‘a line going up’ into ‘a line going up because of a known event.’ Most figures gain a lot from one piece of structural context.”
Already done? Run
ls exhibits/unemployment-rate/figure-v3.png. If it exists,
skip to Step E.
Now the user picks their style. This is the part that matters for the rest of the curriculum: whatever the user decides today is what every future figure starts from.
Walk them through three choices:
figsize=(8, 6)
is 4:3. That’s the right default for slide decks (two figures fit per
slide). 16:9 fills a single full slide. 3:2 is good for memos. Pick one
— 4:3 is the recommended default unless the user knows they want
otherwise.plt.rcParams.update({"font.family": "Times New Roman", "font.size": 12})
at the top of the file.#1f4e79 (deep blue) is fine; if the
user wants their own, set a list. Hex codes only; no named colors.Add plot_v3:
plt.rcParams.update({
"font.family": "Times New Roman",
"font.size": 12,
"axes.spines.top": False,
"axes.spines.right": False,
"axes.grid": False,
})
PALETTE = ["#1f4e79", "#c00000", "#7f7f7f"]
def plot_v3(df, out):
fig, ax = plt.subplots(figsize=(8, 6)) # 4:3
ax.plot(df["date"], df["rate"], color=PALETTE[0], linewidth=1.5)
for start, end in RECESSIONS:
ax.axvspan(pd.Timestamp(start), pd.Timestamp(end), color=PALETTE[2], alpha=0.3)
ax.set_xlabel("Date")
ax.set_ylabel("Unemployment rate (%)")
ax.set_title("U.S. Unemployment Rate")
fig.tight_layout()
fig.savefig(out, dpi=150)Run, open. The figure should now look like a figure the user would put their name on.
Iterate as needed. If the user wants the title removed, the y-axis to start at 0, the line thicker, the recession bar a different shade — do it. Each change is one line of Python; the user is steering, Claude is typing.
Already done? Open
~/ai-training/MEMORY.md and look for a
## Figure preferences section with choices already filled
in. If it’s there, skip to Wrapping up.
This is the durable part. Open ~/ai-training/MEMORY.md
and add — Claude writes, user reviews:
## Figure preferences
- Aspect ratio: 4:3 (figsize=(8, 6)) by default. Two-figure slides fit at this size.
Override to 16:9 (figsize=(10, 5.625)) for full-slide solo figures, 3:2 for memos.
- Font: Times New Roman 12pt. Black labels.
- Palette: ["#1f4e79", "#c00000", "#7f7f7f"] (deep blue primary, red accent, grey for shading).
- Spines: top + right removed. No gridlines.
- Recession shading: NBER dates, grey at alpha 0.3.
- DPI: 150 for saved PNGs.
- Title: short, sentence case, ≤8 words.Tell the user: “Edit any of those to your taste. From now on, when we make a figure, I’ll read this section first. You’re not picking colors twice.”
Name these out loud:
manifest.txt. A figure built on the
wrong CSV is worse than no figure; the verify step is what makes you
trust the result.make_plot.py is
real Python that re-runs in one command. Not one-shot tool-call code
that disappears. Re-running with new data takes 5 seconds.figure-v1.png,
figure-v2.png, figure-v3.png — same logic as
the memo’s versioned drafts. Cheap to keep, expensive to lose.exhibits/ as a canonical folder.
Distinct from exploratory work. Anything ready to ship into a memo or
deck lives here. Other tools (the deck builder in 2.9) will look
here.Three things to try this week:
make_plot.py as a starting point. Notice how
much faster v1 lands now that style preferences exist.Feedback.
The user submits feedback at https://docs.google.com/forms/d/e/1FAIpQLSfBc9jULaJzQWTtAM8dKVCS51JIupf7SVybxCVitWp9aCreyg/viewform.
Claude: paste the URL into chat. The form mirrors the questions below. Collect answers conversationally first, then have the user click through and submit.
Tell the user: “Your instructor uses these to tailor next week’s session.”
Most figures don’t need fancy charting libraries.
Matplotlib does 90% of what people use plotly, seaborn, and ggplot for.
Don’t reach for a heavier library until you’ve hit a wall with
matplotlib’s defaults plus rcParams.
tight_layout() solves most layout
issues. Labels cut off, titles overlapping, weird whitespace —
fig.tight_layout() before savefig fixes most
of it. Reach for bbox_inches="tight" in
savefig if tight_layout isn’t enough.
The data should never be regenerated by hand. Once
data/raw.csv is downloaded and hashed, treat it as
immutable. If the data changes, write
data/raw-2026-05-15.csv and update the manifest. Editing
CSVs by hand is the most common source of figure errors that pass review
and embarrass you later.
Pandas is overkill for tiny data and exactly right for
everything else. A 100-row CSV could be parsed with
csv.reader. Don’t bother — pd.read_csv is the
same line of code and stays right when the data grows.
Excel files work too.
pd.read_excel(path, sheet_name=...) reads
.xlsx directly. For .xlsx files with structure
that matters (multi-sheet, merged cells, Exhibit Two on Sheet 3),
openpyxl lets you treat the workbook as data rather than
copy values manually. Treat the spreadsheet as a source, not a
manual-copy target.