AI-Native Ways of Working
Chapter 2 / Essay
Chapter 2 № 02 · 2026

For code, the skill is not writing.
The skill is using.

Externalize macros, charts, and pivots into Python — the first thing to do

2026.05.02 🌐 日本語

Switch the tool you write logic in to Python.

Tip 5 of the Manual said "what AI is actually best at is, in fact, writing programs." This chapter lands that tip in the actual tooling.

That single change turns repetitive work into one-time work. Reformatting Excel, totaling emails, extracting from PDFs, batch-renaming files. Most "30 minutes of manual office work" finishes in 10 lines of Python. Claude writes it. A human runs it.

Python is for everyone

Drop the prejudice that Python belongs to engineers.

Python is the most readable language AI can write in. Unlike Java or C#, no long class definitions or type annotations are needed. The processing you want to express stands as it is.

import json

with open("orders.json") as f:
    rows = json.load(f)

total = sum(r["qty"] * r["price"] for r in rows)
print(f"Total: {total}")

That is all it takes to read JSON and compute a total. No prior knowledge required. "Open the JSON, multiply qty by price, sum it up" — that is what it says. JSON carries types, so no int() conversion is needed (chapter 4 covers why we choose JSON / SQLite over CSV).

The skill of writing this code is not required. The skill of reading it is enough. If you can read it, you can judge whether the code Claude returned looks right.

Not "writing" — "using"

Here is the new literacy.

Old common sense: learning programming = memorizing language syntax, designing algorithms, writing code.

New common sense: using programming = articulating what you want to process, having Claude write code, running it, checking results.

Compare the time it takes to memorize Excel functions with the time it takes to start having Claude write Python for you. The latter is dramatically shorter. Excel functions only work inside Excel; Python works on any data.

The skill is not writing. The skill is using. This is the new literacy.

You don't need to learn "how to write code." You need to learn "how to put what you want to process into words." That is not technical skill — it is mental clarity.

How to ask Claude

There are only three tricks to having Claude write Python for you.

One: state the input and the output

"Read the Excel file orders.xlsx, total sales per product, and write to JSON summary.json." Input file and output file, both with their formats clear, and Claude does not get lost.

Two: ask one thing at a time

Not "do everything," but "first, write the code to load the data; next, the code to aggregate; last, the code to write JSON / SQLite." Step by step. You can verify behavior partway. When something is wrong, it is easy to back up.

Three: look at the result and correct

The first version is rarely perfect. Run it, look at the output, send back "this is wrong" or "change this." The conversation converges on the right answer. This is what "using" code instead of "writing" it looks like.

What kinds of work become Python

Most everyday tasks for office workers and sole proprietors.

File / data work:

Examples by role:

Work that "a human keeps doing manually" becomes Python. Once written, it works next month and the month after. Chapter 12's "from silos to individual autonomy" — these automations are what technically supports that autonomy.

Have a runtime

To use Python, you need a runtime.

Three options.

One: install Python on your own machine

Run with the python command. Files on your machine just work. The first step has a small hurdle, but once installed, it works forever. Mac and Linux often have Python preinstalled. On Windows, install from the Microsoft Store.

For package and tool installation, default to uv (a fast Rust-built installer). For data science and scientific computing workflows where uv struggles — numpy / scipy / pytorch / tensorflow / GPU libraries / GIS / bioinformatics with heavy C/C++/Fortran dependencies — switch to Miniforge (the FLOSS conda distribution that uses conda-forge by default; not bound by Anaconda's commercial terms). Both are equally usable from AI — ask Claude "install with uv" or "install with conda" and the right commands come back. The "skill of using, not writing" applies to package managers too.

Two: use Claude's code execution

Paste code into Claude and ask "run this." For light processing, you don't need to set up an environment locally.

Three: use Google Colab

Google's browser-based Python service. Free. Suited to heavier data processing.

If unsure, just start. Ask Claude "I want to set up a Python environment. Ask me what suits me, and tell me." It will guide you based on your situation.

Don't fear "it didn't run"

In the early days, you'll often see errors when running Python.

That is normal. Copy the error text into Claude and say "I got this." Claude will identify the cause and return corrected code. An error is not the end; it is input for the next instruction.

Paste the error text into Claude. Claude fixes it. You move on.

You don't need to feel "programming isn't for me." If you can paste an error message, that is enough.

This is exactly the "trust the handoff, check the result properly" practice from the prologue — let Claude write the code, run it, look at the result, and ask in words for fixes. Whether you wrote the code yourself or Claude did, the testing practice is the same.

Readable in ten years

Python has been around for over thirty years. Some code broke in the migration from Python 2 to 3, but Python 3 code will keep running for another 10 or 20 years.

Excel VBA macros may stop working with each new Office version. Python code is text, dependencies are explicit, and AI can re-read it.

Hold logic, too, as structure.

VBA runs today's Excel. Python keeps running across time.

In numbers

Monthly job extracting amounts from 100 invoice PDFs: by hand, 4 hours. With Python written by Claude, 3 seconds. Next month, the same script in 3 seconds. A 4,800x gap on the first run; infinite afterward.

Office work aggregating 50 Excel files monthly: 5 minutes per file × 50 = 4 hours. Write the pandas Python script once, and python aggregate.py runs in 10 seconds thereafter.

Python learning curve: for the "ability to use," practicing reading Claude-written code for one week is enough to start using at work. The traditional self-study path to "ability to write" takes 6 months. A 24x gap.

Reaction to errors: searching and trial-and-error solo takes 30 minutes to 2 hours. Pasting the error text into Claude returns the cause and corrected code in 30 seconds. 60x faster or more.

In summary

Align your tools with the AI era, and the way you process work changes.

From Excel manual labor to Python plus Claude. A single step turning repetitive work into one-time work. The skill of writing is not required. The skill of using is enough.

The next chapter moves on to writing documents — from Word to Markdown. With Python tooling now in place, the substance of documents can also move from formatting to structure.

Related

Examples

Runnable source, commands, and measured results — see the dedicated example page(s).