Two Years Ago, My Day Looked Completely Different
Believe it or not, two years ago, I was still writing and debugging code daily—line by line. I spent hours slamming my head against the desk, only to realize the bug was a typo. It sounds crazy, right?
A typical day then involved:
- Writing and debugging every SQL query and Python script from scratch.
- Building slide decks bullet by bullet, manually formatting every chart.
- Writing documentation that no one would read until something broke—and even then, they'd barely glance at it.
I published a day-in-the-life piece in 2024, but almost none of that describes my actual Tuesday in 2026.
I won't pretend that's purely a good thing. Some days it feels less like my job got easier and more like it quietly morphed into a different role—one I had to learn on the fly while still doing the old one.
(And yes, plenty of data scientists still do many of the things I did two years ago, but this reflects my experience and that of many colleagues I know today.)
Prompt Engineering Is a Core Skill
Sure, we had ChatGPT in 2024, and we knew about prompt engineering. But I avoided it—ChatGPT often frustrated me. Explaining the context before feeding it code was more work than just fixing the bug myself, and even then, it often missed the root cause.
That has changed dramatically with newer AI models, especially Claude. Features like projects and skills allow the AI to understand the context and history behind a project, making collaboration much smoother.
So a significant chunk of my day now goes into writing and refining prompts. Early on, my prompts were lazy:
Summarize the forecast accuracy for this model.
That produced a vague paragraph that often missed the insights I needed. Now, I write prompts like:
Summarize this model's forecast accuracy over the last 14 days. Report the exact MAPE and RMSE for each day, flag any day where MAPE exceeded 5%, and state whether the trend is improving or degrading week-over-week. Do not round error metrics; report them to two decimal places.
The difference in output quality is enormous. Honing this skill is now a daily practice.
Key parts of my workflow include:
- Double-checking LLM model outputs for accuracy and consistency.
- Testing prompt variants against the same task, comparing outputs side by side.
- Embedding constraints directly into prompts (units, decimal precision, what not to guess at) rather than correcting outputs afterward.
Finding Cost-Effective LLM Solutions & Cutting Token Usage
LLMs are expensive—far more so than traditional models like XGBoost. This means we must be more thoughtful about using LLMs to analyze large datasets.
The same data science principles still apply, but now with an added layer of cost optimization. I've implemented several strategies to keep expenses in check:
- Set token budgets: For each task, I define a maximum token limit and monitor usage in real-time.
- Use cheaper models for simple tasks: Not every query needs a top-tier model; for straightforward SQL checks, smaller models work fine.
- Cache common queries: Repeated analysis on daily data can be cached, reducing redundant LLM calls.
- Batch processing: Instead of sending many small requests, I batch them into fewer, larger ones to reduce overhead.
These practices have cut our LLM spend by over 30% without sacrificing insight quality.
From Code Monkey to AI Orchestrator
The biggest shift in my role is that I'm no longer just writing code; I'm orchestrating AI agents that write and execute code for me. I spend more time defining the problem and evaluating outputs than I do debugging syntax errors.
For example, I now have a set of skills in Claude that encode our team's best practices—how to handle missing data, preferred chart styles, and common error messages. When I start a new analysis, the AI already knows our standards.
This has freed up time for higher-level thinking, but it also comes with new challenges:
- Oversight is critical: I can't blindly trust outputs; I must validate results and understand the reasoning.
- Bias and hallucination risks: AI can generate plausible but wrong conclusions, so critical thinking is more important than ever.
- Staying current: The AI landscape evolves monthly, so I dedicate time each week to learning new features and best practices.
A Realistic Tuesday in 2026
So, what does my Tuesday actually look like?
- 9:00 AM: Review overnight LLM-generated summaries of key metrics and flag anomalies.
- 10:00 AM: Refine prompt templates for a new client dashboard, testing variations against historical data.
- 11:30 AM: Meet with engineering about integrating our AI models into production pipelines.
- 1:30 PM: Deep-dive into a model performance issue—using the AI to suggest feature engineering ideas, then validating them.
- 3:00 PM: Write and deploy a new prompt workflow for anomaly detection, using cost-saving strategies.
- 4:30 PM: Team stand-up where we discuss lessons learned and share prompt best practices.
It's less about writing code from scratch and more about guiding AI to do it efficiently and correctly.
The Human Element Remains
Despite all the automation, the core of data science is unchanged: asking the right questions, understanding the business context, and communicating insights clearly. AI has removed the drudgery, but it has amplified the importance of judgment.
In 2026, a data scientist who can't write a precise prompt or manage token budgets is at a severe disadvantage. But those who embrace the shift are finding their work more impactful—and more human.
Let me know in the comments how your role has changed—or not—in the age of AI. I'd love to hear your perspective.