How to use the Credit Risk Data Cleaning agent skill

What is an agent skill and what does this one do?

An agent skill is a folder of instructions and scripts that an AI agent can load when a task matches the skill's description. This one contains a SKILL.md file, two Python modules in references/ (analysis.py and func.py), and an example script in scripts/example.py.

Once installed, the agent can run an 11-step cleaning pipeline. Each step runs independently and does not delete the original data, so results can be compared between steps.

• Load and format the raw data
• Count samples and bad sample rate per organization, and separate out-of-sample (OOS) data
• Filter abnormal months with too few bad or total samples
• Calculate missing rates and drop features that are mostly empty
• Drop features with low IV (information value) overall or in too many organizations
• Drop features with unstable PSI (population stability index)
• Remove noise features with Null Importance denoising
• Remove highly correlated features based on original gain
• Export an Excel cleaning report

How do you install the skill?

Download the ZIP file for datanalysis-credit-risk from this page. Then choose one of the two options below.

For a personal install across all projects, unzip the file into ~/.claude/skills/. For a single project, unzip it into that project's .claude/skills/ folder. You can also upload the skill in the Claude apps under Settings → Capabilities → Skills.

• Personal: ~/.claude/skills/
• Project: .claude/skills/ inside your project folder
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

What do you need before running it?

The skill runs Python scripts, so Python must be available. It also needs several Python packages and a credit dataset in one of the supported formats. The source lists parquet as the best format for DATA_PATH, and also accepts csv, xlsx or pkl files.

• Python
• pandas, numpy, toad, lightgbm, scikit-learn, openpyxl, joblib, tqdm
• A credit dataset in parquet, csv, xlsx or pkl format
• Columns for the date, label (Y), organization and primary key

How do you trigger the skill with a natural request?

You do not need to type a command. Describe the task in plain language and mention credit data cleaning or variable screening. The agent matches your request to the skill's description and loads it. Name your file and columns so the agent can set the parameters.

Example requests you can type:

• Clean my credit dataset at data/loans.parquet. The date column is apply_month, the label is y, the organization column is org and the key is loan_id.
• Screen the variables in this loan data before I build a pre-loan model, and give me an Excel report.
• Remove features with more than 60% missing values and low IV from this credit file, then tell me what was dropped.

What does a run look like?

The quick-start script is run with Python from the skill folder. The skill's own instructions show this command:

• Parameters can be entered before each step, and default values are used if you leave them blank.
• Defaults include a minimum of 10 bad samples and 500 total samples per month, a missing rate threshold of 0.6, an IV threshold of 0.1 and a PSI threshold of 0.1.
• IV and PSI calculations can use multiple processes to run faster.
• The final output is an Excel workbook with one sheet per step, plus a summary sheet.

What should you check before using it on sensitive data?

The skill runs scripts, so read the files before running them on your machine. Run it locally on data you are allowed to process.

The source states that the skill does not use the network. It also notes that network use is not documented in the files, so check this yourself before running it on sensitive data that contains personal information. The skill only filters data and does not validate a model, so model validation and human review are still needed before any lending decision.

• Runs scripts: yes
• Uses network: no (as stated in the listing, which also advises checking before use on sensitive data)
• The skill cleans and screens data; it does not approve or reject loans
• Keep your original data backed up, although the pipeline does not delete it

Tips for getting reliable results

Check the defaults against your own policy before running. The thresholds are parameters, so adjust them rather than accepting them blindly.

Read the Excel sheets for the steps that matter most to you, especially the removed-feature sheets, to confirm each drop makes sense.

• Convert your data to parquet where possible, since the source recommends it.
• Set the organization and OOS organization lists correctly so OOS samples are separated from modeling samples.
• Use the step-by-step sheets (for example Step7-IV明细 and Step8-PSI明细) to trace why a variable was removed.
• Record the parameters you used so the cleaning can be repeated.