Vibe coding · prompt builder 仕様

Build the setup prompt

The first message of a project carries the rules and asks for no code. Fill in what you know. Whatever you leave blank stays in the prompt as a marked gap, so you can see what the assistant would have had to guess.

The task What you are building Completes the sentence “You will help me build …”. Kind of problem Sets how the data is split, the default score and the baseline.

Rule 01 · Environment

Where the code runs

Google Colab Browser notebook with an assistant built in Laptop · Jupyter An .ipynb file in VS Code or JupyterLab Laptop · script .py files run from the terminal

Project folder

Rule 02 · Data contract Source

Rows Columns

Target column Its values

Columns the assistant would misread One per line: codes that look like numbers, dates stored as text, columns filled in after the outcome.

Rule 03 · Model and search space Models allowed Search space One parameter per line. Leave it empty if you tune in a later step.

Off the table

Any other model, AutoML included Feature scaling Resampling (SMOTE, over- or undersampling) Dimensionality reduction (PCA)

Random seed

Rules 04–05 · Validation

Test set, % CV folds

Score to tune on A scikit-learn scoring name. It changes with the kind of problem until you edit it. Threshold rule Optional. The prompt already says the threshold comes from out-of-fold predictions.

Rule 06 · Success The goal, in business terms Metrics to report

Baseline to beat: the majority-class baseline (predict the most common class for every row).

Rules 07–08 · Working style

One step per answer Explain before each code block Print every number, invent none Ask before adding a library List the files touched Smallest fix when a rule breaks

Prompt 0 · Setup

ROLE AND GOAL
You will help me build <what you are building, in one sentence>, step by step. Do not write any code in this answer.

ENVIRONMENT
- Google Colab notebook, Python 3. pandas, scikit-learn and matplotlib are already installed.
- Give the code as separate notebook cells. Put a short markdown cell (a title and 1–2 lines) before each code cell.
- Install extra packages only if essential, with !pip install in the first cell.
- Read the data directly from the source under DATA. No Google Drive, no local paths.
- Work inside /content/my-project with folders data/raw, data/processed, outputs/figures, outputs/models and outputs/reports (create them with pathlib).
- Show plots inline.

DATA
- Source: <URL or path>. <rows> rows, <columns> columns.
- Target: "<column>".

MODEL SETUP (hard constraints)
- Use ONLY <the model classes you allow>.
- No other models (AutoML included), no feature scaling, no resampling (SMOTE, over- or undersampling), no dimensionality reduction (PCA).
- random_state=42 everywhere.

VALIDATION RULES
- Split first: stratified 80/20 train/test with random_state=42, before any preprocessing.
- Fit every transformation (imputers, encoders, scalers) on train only.
- Tune hyperparameters only with 5-fold StratifiedKFold cross-validation (shuffle=True, random_state=42) on train, scoring="roc_auc".
- Choose the decision threshold on out-of-fold predictions, never on test.
- Use the test set exactly once, at the end.

SUCCESS
- Goal: <what a good result lets someone do>.
- Report recall and precision for the positive class, ROC AUC and the confusion matrix.
- Always compare with the majority-class baseline (predict the most common class for every row).

WORKING STYLE
- One step per answer; wait for me to say "next".
- Before each code block, 2–3 lines on what you will do and why.
- Print every key number. Never invent results.
- Ask before adding any library.
- In every answer, list the files you created or changed.
- If I say a rule is broken, redo only the affected code and keep everything else.

Reply only with: (1) the task in 5 bullets, (2) your step plan, (3) any question.

7 blanks left to fill.

The form needs JavaScript to update the prompt. Without it, the text above is the template: replace everything in angle brackets by hand.

What you type stays in this browser. Nothing is sent anywhere.

How to send it

First message, no code

Send the setup prompt on its own. The reply should restate the constraints, propose a step plan and ask its questions. If it comes back with code, the working-style rules did not land. Say so before going further.

Then one step per message

Load and audit the data. Split and set the baseline. Tune inside train. Open the test set once. Explain the result. Each step is a short prompt of its own, and each answer gets checked before the next one goes out.

Same prompt, another tool

To move from Colab to a laptop, change the environment and nothing else. If two tools give very different numbers from the same spec, one of them broke a rule. The checklist will tell you which.

Build the setup prompt

How to send it

Prompts live in a folder

steps/00_setup.md