When Data Scientists Become Prompt Engineers
Data Sci
Code for production. Prompt for exploration. Know which is which.
When Data Scientists Become Prompt Engineers
TL;DR
- LLMs can generate code, explain results, and draft analyses. For exploration and prototyping, that's fast.
- For production, reproducibility, and audit trails — you still need real code.
- The shift: less "write Python from scratch," more "direct the AI, validate the output, own the pipeline."
It's uncomfortable. You spent years learning stats, Python, SQL. Now you can get a working analysis by chatting with a model. Some of that work is commoditized. The parts that remain valuable: asking the right question, validating the answer, and putting it in a form the business can trust.
What LLMs Are Good At (For Data Work)
Code generation:
- "Write a pandas pipeline to aggregate X by Y" — Works. Saves time.
- "Debug this error" — Often works. Copy-paste, run, iterate.
Exploratory analysis:
- "What might be causing this correlation?" — LLM suggests hypotheses. You test them.
- "Explain this chart" — LLM drafts narrative. You fact-check.
Documentation and reporting:
- "Summarize these findings for stakeholders" — First draft in seconds. You edit.
- "Create a README for this analysis" — Boilerplate done. You customize.
Learning and reference:
- "How does X work in library Y?" — Faster than docs sometimes. Verify before using.
What You Still Own
Problem framing:
- "What should we predict?" "What's the right metric?" — LLMs can suggest. You decide. The business context is yours.
Validation:
- LLMs hallucinate. Code has bugs. Analyses have flaws. You're the skeptic.
- Run the code. Check the numbers. Sanity-check the conclusions.
Reproducibility:
- A prompt session isn't a pipeline. Production needs versioned code, dependencies, and audit trails.
- LLM-generated code → you refactor, test, commit.
Interpretation and storytelling:
- "What does this mean for the business?" — LLMs can draft. The nuance, the caveats, the "here's what we don't know" — that's you.
The New Workflow
- Prompt for exploration — Fast iteration. Get to "does this make sense?" quickly.
- Code for production — Refactor LLM output. Add tests. Version. Deploy.
- Validate everything — Don't trust. Verify. Especially when the answer looks right.
Write Python from scratch. Debug. Document. Reproducible pipeline. Days per analysis.
Click "Data Work With LLM Assistance" to see the difference →
Quick Check
An LLM generated analysis code that produces a surprising result. What do you do?
Do This Next
- Run one analysis twice — Once with pure coding. Once with LLM assistance. Compare time and quality. Where does the LLM help? Where does it hurt?
- Define your "prompt vs. code" boundary — What's exploration (prompt OK)? What's production (code required)? Write it down.
- Build a validation checklist — For any LLM-assisted analysis: What do you always verify? Add it to your process.