CSV audit and production-ready cleaning pipeline

Audits a raw dataset for data smells, ties each issue to its business impact, then proposes an imputation, encoding and scaling strategy with a leakage-safe pandas and scikit-learn pipeline.

PLPromptByte Library@promptbytePublished Aug 29

TextData & Analysis #data cleaning #data quality #etl #pandas #python

Version 1 of 1 · published Aug 29, 2026Version 1 · only version

Example output

See all results

Anthropic · Claude Opus 5 PLPromptByte Library Version 1

Prompt

Version 1 · 3 variables · 2 required

Fill in the values below — the customized prompt updates as you type. marks a required value.

Dataset description{{dataset_description}}

Rows, columns, dtypes, and any known quality problems. Pasting df.info() output works well.

Business objective{{business_objective}}

The decision or product this cleaned data has to support.

Target variable(optional){{target_variable}}

The column being predicted, if any. Drives the data-leakage checks. Default: none

Fill in Dataset description, Business objective to copy the customized prompt.

Resultsfor version 1

All models Claude Opus 51

Version 1 · only version

Claude Opus 5 result by PromptByte Library
Anthropic · Claude Opus 5 PLPromptByte Library (@promptbyte)Version 1Published Aug 29
Compare
## 1. Technical Audit & Business Context | Column | Smell | Business impact | |---|---|---| | `churn_date` | Populated only for churned accounts | This is the label in disguise. Any model that sees it will score ~perfectly in training and be useless in production. | | `mrr` | Right-skewed, a long tail of enterprise accounts | Mean imputation would drag small accounts upward and understate churn risk in the SMB segment, which is where outreach capacity actually sits. | | `signup_date` | Mixed `YYYY-MM-DD` and `DD/MM/YYYY` strings | Tenure is derived from this. Silent misparsing flips day and…
Community submittedFull output
Settings: Unknown
Open result

Showing all 1 public result from Claude Opus 5 on version 1.

Usage notes

Describe the dataset concretely in dataset_description: row count, column count, the column names and types that matter, and anything you already know is messy. Pasting df.info() and df.describe() output works very well here.

Naming a real business_objective is what makes this prompt better than a generic "clean my data" request — the audit step will rank issues by how much they distort that specific decision.

If you are preparing data for a model, always fill in target_variable; that is what drives the leakage checks and the fit-on-train-only handling inside the Pipeline. Leave it as none for pure reporting or BI work.

Treat the generated script as a reviewed starting point. Run it on a copy, check the assertion block actually fires on bad input, and confirm the imputation choices against your domain knowledge before scheduling it.

Attribution: adapted from https://github.com/f/prompts.chat by f (Fatih Kadir Akın), licensed CC0-1.0. Source file: https://github.com/f/prompts.chat/blob/main/PROMPTS.md

Output2,714 characters

## 1. Technical Audit & Business Context | Column | Smell | Business impact | |---|---|---| | `churn_date` | Populated only for churned accounts | This is the label in disguise. Any model that sees it will score ~perfectly in training and be useless in production. | | `mrr` | Right-skewed, a long tail of enterprise accounts | Mean imputation would drag small accounts upward and understate churn risk in the SMB segment, which is where outreach capacity actually sits. | | `signup_date` | Mixed `YYYY-MM-DD` and `DD/MM/YYYY` strings | Tenure is derived from this. Silent misparsing flips day and month for the first twelve days of each month and corrupts every cohort chart. | | `country` | Free text, with `US`, `USA` and `United States` all present | Splits one segment into three, so no country clears a significance threshold. | | 12 usage counters | Zero-inflated; nulls appear to mean "no events", not "unknown" | Imputing the median here invents activity for dormant accounts, which is exactly the signal you are trying to detect. | ## 2. Statistical Strategy - **Imputation** — median for `mrr` and tenure (robust to the enterprise tail); constant `0` for the usage counters, since missing means no events; a `Missing` category for `country`. - **Encoding** — one-hot for `plan` (3 levels, no order) and for a normalised `country_region` (5 levels after cleanup). No label encoding: it would impose a false ordering that tree splits and linear coefficients would both read as real. - **Scaling** — RobustScaler on `mrr` and the counters. StandardScaler assumes roughly symmetric spread and would compress the informative high-usage tail. ## 3. Implementation Block ```python numeric = ["mrr", "tenure_days"] counters = [c for c in df.columns if c.startswith("usage_")] categorical = ["plan", "country_region"] preprocessor = ColumnTransformer([ ("num", Pipeline([("imp", SimpleImputer(strategy="median")), ("sc", RobustScaler())]), numeric), ("cnt", SimpleImputer(strategy="constant", fill_value=0), counters), ("cat", Pipeline([("imp", SimpleImputer(strategy="constant", fill_value="Missing")), ("ohe", OneHotEncoder(handle_unknown="ignore"))]), categorical), ]) ``` Fit the ColumnTransformer inside a Pipeline on the training split only; that is what keeps the imputation medians and scaler quantiles from leaking test-set information. ## 4. Post-Processing Validation ```python assert "churn_date" not in X.columns, "label leakage: drop churn_date" assert X.isna().sum().sum() == 0, "nulls survived the pipeline" assert X.select_dtypes("float64").empty, "downcast float64 -> float32" ```

code · sql
```
**Assumptions**

- `order_items.order_id` joins to `orders.id`, and `products.category_id` joins to `categories.id`. These are inferred from the naming convention, since no foreign keys were listed.
- "Last 90 days" is measured against `orders.order_date` relative to today, inclusive of today.
- Revenue is `quantity * unit_price` summed over line items; there is no discount or tax column in the schema, so none is applied.
- Cancelled orders are excluded. If `status` uses different values than `'
```
Code preview: **Assumptions** - `order_items.order_id` joins to `orders.id`, and `products.category_id` joins to `categories.id`. These are inferred from the naming convention, since no foreign keys were listed. - "Last 90 days" is measured against `orders.order_date` relative to today, inclusive of today. - Revenue is `quantity * unit_price` summed over line items; there is no discount or tax column in the schema, so none is applied. - Cancelled orders are excluded. If `status` uses different values than `'
SQL query generator from a plain-English request
Converts a natural-language data requirement plus a table schema into a syntactically correct SQL query for a named database, with its assumptions stated up front.
CodeData & Analysis OpenAI · GPT-5.5
PL@promptbyteAug 27
Score 1
0 comments1 results
No public results yet
Audit SQL joins for duplicate rows and inflated totals
Check join cardinality and aggregation grain to explain duplicate rows or double-counted measures before changing a reporting query.
TextData & Analysis
PL@promptbyteSep 8
Score 0
0 comments0 results
No public results yet
Design a forecast backtest that respects time and information
Specify temporal evaluation, leakage checks, baselines and business-relevant metrics before judging a forecasting model.
TextData & Analysis
PL@promptbyteSep 8
Score 0
0 comments0 results
No public results yet
Critique an existing chart for misleading interpretation
Review a chart's scales, labels, aggregation and claims, separating visual issues from data questions that need verification.
TextData & Analysis
PL@promptbyteSep 8
Score 0
0 comments0 results
No public results yet
Classify comment sentiment with traceable evidence
Label supplied comments consistently while preserving mixed opinions, unclear cases and the exact words supporting each label.
TextData & Analysis
PL@promptbyteSep 8
Score 0
0 comments0 results
No public results yet
Reshape a wide CSV into tidy records without losing meaning
Produce a deterministic transformation from repeated measure columns to tidy rows with explicit keys, types and missing-value rules.
CodeData & Analysis
PL@promptbyteSep 8
Score 0
0 comments0 results

Browse Data & Analysis

CSV audit and production-ready cleaning pipeline

Example output

Prompt

Resultsfor version 1

Models represented

Something wrong with this prompt?

SQL query generator from a plain-English request

Audit SQL joins for duplicate rows and inflated totals

Design a forecast backtest that respects time and information

Critique an existing chart for misleading interpretation

Classify comment sentiment with traceable evidence

Reshape a wide CSV into tidy records without losing meaning

Discussion 0

Example output

Prompt

Resultsfor version 1

Models represented

Something wrong with this prompt?

Related prompts

SQL query generator from a plain-English request

Audit SQL joins for duplicate rows and inflated totals

Design a forecast backtest that respects time and information

Critique an existing chart for misleading interpretation

Classify comment sentiment with traceable evidence

Reshape a wide CSV into tidy records without losing meaning

Discussion 0