Learn to GPT

Learn to GPT

For Analysts & Researchers

Data work splits cleanly between Claude and ChatGPT

One executes code, the other reads everything at once

ChatGPT's data analysis mode actually runs Python on your uploaded files, which makes it the safer choice for anything arithmetic. Claude's edge is the other half of the job: reading a huge pile of documents or feedback in one conversation, extracting it into clean structure, and writing the narrative on top. This page maps which task goes to which tool.

Start Learning Free Prompt Engineering

The Split

Three rules for dividing the work

If it needs computing, use executed code

A language model predicting the sum of a column is guessing, however confidently. ChatGPT's data analysis mode sidesteps this by writing Python and running it on your actual file. Claude's analysis tool can run JavaScript in-browser for similar checks. Either way: for numbers that matter, insist on executed code you can read, not model arithmetic.

If it needs reading, use the biggest context

Synthesis quality drops when a tool has to chunk your sources and summarize the summaries. Claude's context window is large enough to hold a stack of papers, a long annual report, or months of support tickets in one conversation, which is why document-heavy analysts tend to keep it open next to ChatGPT rather than instead of it.

If it needs judgment, any of them can push back

Correlation versus causation, sampling bias, a question framed so the answer flatters the asker: the frontier models are all capable of catching these when you ask them to play skeptic. That prompt ("what would a hostile reviewer say about this analysis?") transfers across Claude, ChatGPT, and Gemini unchanged.

Workflows

Six data tasks, routed to the right tool

Spreadsheets: run the numbers in ChatGPT

For real computation (sums, pivots, regressions) upload the file to ChatGPT's data analysis mode, which writes and executes Python. Language models doing mental math over pasted CSVs make mistakes; executed code doesn't. Use Claude when the question is interpretive rather than arithmetic.

Upload sales.csv, then: run a month-over-month growth calculation per product line in Python, show the code, and chart the three fastest-growing lines. Flag any month where the data looks incomplete before analyzing it.

SQL generation: a genuine tie

Describe your tables and the question; any frontier model writes solid Postgres, MySQL, or BigQuery. Paste your actual schema DDL and sample rows to cut hallucinated column names. Whichever tool you use, read the query before running it against production.

Given these CREATE TABLE statements and three sample rows per table, write one query answering: which signup cohorts have the highest 90-day repeat-purchase rate? Explain each join in a comment.

Research synthesis: Claude's home turf

This is the task that justifies the second tab. Drop several papers or reports into one Claude conversation and interrogate the whole pile at once. Its long context holds more source material in view than most alternatives, which changes what questions you can ask.

You have the full text of several studies on this topic. Build a claims matrix: each row a finding, columns for which studies support it, dispute it, or don't address it. Then tell me where the literature genuinely disagrees and why.

Structured extraction: Claude's other strength

Turning messy text (emails, postings, transcripts) into JSON or CSV rows against a schema you define. Claude has a reputation for holding a schema faithfully across many documents; GPT models do this well too with structured-output mode. Spot-check either.

Convert each of these vendor emails into one JSON record: {vendor, product, quoted_price or null, delivery_weeks or null, open_questions[]}. Where the email is ambiguous, put your uncertainty in open_questions rather than guessing.

Report narratives: pick by writing taste

Turning finished numbers into an executive summary is a writing job, not a math job. Many people prefer Claude's prose defaults here, but this is preference, not capability. Give whichever model you use the audience, the length cap, and one example of a past report you liked.

Using these final Q2 figures (already verified), draft the two-paragraph summary for a board pack: headline result first, one risk named plainly, no adjectives doing the work numbers should do. Match the attached example's tone.

Framing the analysis: use either as a thought partner

Before any tool touches the data, talk through what's worth asking. Describe the columns and the business question, and let the model propose analyses, confounds, and cheap first checks. This works equally well in Claude, ChatGPT, or Gemini.

I have 18 months of support tickets with timestamps, categories, resolution times, and satisfaction scores. Before I compute anything: what confounds should worry me, what would a skeptic check first, and what result would actually change a staffing decision?

Code Examples

The same split, from the API side

Python: send a data preview, not the whole file

# Same pattern works with the OpenAI SDK;
# swap the client and model name.
import anthropic
import pandas as pd

df = pd.read_csv("sales_data.csv")
preview = df.head(20).to_csv()
dtypes = df.dtypes.to_string()

client = anthropic.Anthropic()
response = client.messages.create(
    model="claude-opus-4-5",
    max_tokens=2000,
    messages=[{
        "role": "user",
        "content": f"""Column types:
{dtypes}

First 20 rows:
{preview}

Suggest the three analyses most worth
running on this, and what each would
tell us. Do not compute anything yet."""
    }]
)
print(response.content[0].text)

Structured extraction against a fixed schema

# Keep the schema in the prompt verbatim
# so every record comes back uniform.
response = client.messages.create(
    model="claude-opus-4-5",
    max_tokens=1000,
    messages=[{
        "role": "user",
        "content": """Return this posting as JSON:
{
  "role_title": string,
  "company": string,
  "salary_min": number | null,
  "salary_max": number | null,
  "remote_policy": "remote"|"hybrid"|"onsite",
  "tech_stack": string[],
  "years_experience": number | null
}
Use null, never guess missing fields.

Posting:
""" + job_posting_text
    }]
)
import json
data = json.loads(response.content[0].text)

By Role