Instead of treating vibe coding as simply 'tell AI what you want and it builds everything for you,' this article frames it as a workflow in which a person defines the requirements and verification criteria, then checks the result produced by AI. Using a fictional household-expense CSV summation tool, it organizes the input, output, things not to do, and verification method into a one-page note.
Who this is forGeneral users with limited coding experience who want to start safely with small automations using a terminal-based AI coding tool such as Claude Code
In this series, vibe coding means describing the desired task in natural language, letting AI help read files or create and modify code, and then having a person verify the result. The key is not a promise that the person will never need to write a single line of code. Instead, the person should first decide what to build, what the input and output are, what is allowed, and how to verify whether the result is correct. As covered in the introductory generative AI article, “What Is Generative AI? Understand How Answers Are Generated and Why Hallucinations Happen,” AI answers can look plausible while still conflicting with facts or intent, so it is safer to separate 'generation' from 'verification' in coding as well.
| Role | What AI can handle | What the person should decide or verify |
|---|---|---|
| Understand requirements | Organize the scope based on the request | Check whether purpose, input, output, or prohibitions are missing |
| Write code | Draft the necessary files and code | Check for unwanted file changes or scope expansion |
| Prepare to run | Suggest commands or the required sequence | Check how the command affects the current folder and files |
| Verify results | Suggest how to compare against the expected result | Directly check whether the actual output matches the reference values |
In other words, the quality of vibe coding is not determined only by how elegantly you write a prompt. It is more important to define the task narrowly, make AI's work easy to inspect, and leave a clear point of reference for where to look again if something is wrong.
If you immediately ask AI to 'make me an expense tracker,' AI may fill in the file format, output style, library choices, and error-handling scope on its own. Even for a small tool, writing down four items first makes the later conversation much clearer: first, what file is the input; second, what values the output should contain; third, what must not be done; and fourth, what the person will compare the result against.
Tool name: Monthly totals for expense CSV
Input
- UTF-8 CSV file expenses.csv
- Columns: date, category, amount
- date format: YYYY-MM-DD
- amount is an integer
Output
- Display each monthly total with its month
- 2026-09 = 26600
- 2026-10 = 9800
Do not
- Use real personal information or real payment data
- Add external packages
- Modify the input CSV
How to verify
- Run python summarize.py expenses.csv
- Check that the two monthly totals in the output exactly match the expected values aboveThis note is not code. However, it becomes the standard for judging whether a proposed implementation goes beyond the requirements. For example, if AI suggests installing pandas, that conflicts with the condition 'use only the Python standard library,' so you do not have to proceed as suggested.
When you hand the task to Claude Code in the next step, you can provide the requirements note as-is or summarize its essentials. A good request does not end with 'make this.' It also includes an input example, expected output, constraints, and what to check before editing. That leaves the person with evidence for deciding whether the generated code is correct.
Make a plan for creating a Python script that reads expenses.csv and prints monthly amount totals.
Do not modify any files yet.
The input columns are date, category, amount; date uses YYYY-MM-DD and amount is an integer.
Use only the Python standard library.
The expected values are 2026-09 = 26600 and 2026-10 = 9800.
First explain which files you will read, what you will check, and the order in which you would implement the task.This request does not yet approve writing code. Adding 'understand and plan first' creates a point where you can verify the scope. Claude Code's official documentation also recommends breaking complex work into steps and having Claude understand the code before making changes.
Because AI coding tools can read and modify files or suggest running commands, beginners should focus less on 'when to press Yes' and more on 'when to stop and inspect.' In particular, if you see an unfamiliar command, a proposal to change many files at once, an unrequested package installation, or code that modifies the original input, check the reason before moving on.
Existing Worknote articles cover code-review checklists and unit testing separately, so those procedures are not repeated here. This series focuses on the workflow in which a person controls task scope and the timing of changes while collaborating with a terminal-based AI coding tool.
After finishing this article, you still do not need to install Claude Code or generate code. Instead, you should have a one-page requirements note. In the next article, you will take this note, install Claude Code, and start by asking it to read and explain the practice folder before changing any code.
| Check item | Criterion for this example | Completion criterion |
|---|---|---|
| Input | expenses.csv / date, category, amount | The file and column names are explicit |
| Output | 2026-09 = 26600, 2026-10 = 9800 | Expected values are fixed numerically |
| Constraints | Use only the standard library; do not modify the original CSV | It is harder for AI to expand the scope arbitrarily |
| Verification | Run python summarize.py expenses.csv and compare with the expected values | A person can directly judge whether the result is correct |
With these criteria in place, the core questions stay simple even if AI later gives a long code explanation: 'Does it read the input I defined? Does it produce the output I defined? Does it avoid the things I prohibited? And can I verify the result?' Check them in that order.
Written using the Claude Code official documentation (checked 2026-09-22) and the fictional example in this series · Claude Code not actually run · No Python code to execute
Claude Code or Python code was not actually run in this article. This step explains the vibe-coding workflow and fixes the requirements for the later hands-on exercises; concrete implementation and execution verification are covered in subsequent articles.
Choose a question that interests you and create a real output.
If you want to understand the principles first: What Is Generative AI? Understand How Answers Are Generated and Why Hallucinations Happen
Build a request that clearly includes your goal and constraints.
Practice with hints and review of wrong answers.
Create a weekly plan that respects your time limits.
Learn how to compare a summary against the original source.
Compare repetitive tasks and pick your team's first use case.