Build a credit risk scorecard in Excel with Claude in Excel
- Posted
- by @akshen121
- Last checked
- Method:
- How we verify
Evidence: Shown on videoThe source includes a video or screen recording of the agent doing the task.
PartialThe source says it only partly worked, or had a significant problem a person had to deal with.
What went wrong: Looked impressive but broke the brief.
Outcome
Given a credit risk dataset and a brief for an Excel-native scorecard, Claude in Excel built a polished 28-sheet workbook but skipped the variable clustering, ran the model in JavaScript and reported 0.706 AUC on training data with no split; ChatGPT followed the method and got 0.627 on test data.
Agent
Prompt Improved prompt
Strengthened by us to avoid the problem in the source. Not tested by the poster. It doesn't change what happened in the source.
Run this task now, exactly as written below. Everything you need is here, so start straight away and only stop if something required is missing. Goal: Build an end-to-end credit risk scorecard inside Excel, following the methodology exactly. Fill in before running: Which dataset and target column, and must everything be Excel-native? What I'll give you: - The dataset ([FILE]) and target (bad/good) column - The required steps Rules: - Follow every step in order: bin variables, calculate Information Value, cluster variables, select features, build the model, plot ROC. - Do all calculations in Excel formulas; don't paste values from code. - Split train/test first and report AUC on the test set only. - If you skip or change a step, say so clearly. - Merge sparse bins so Information Value isn't inflated. - Document each variable's points and why, on a separate sheet. - Test the scorecard on 10 sample rows and show the scores. Output: The workbook, plus a one-page summary of each step and the test AUC. You'll know it worked when: each step is visible in Excel and the AUC is from held-out data.
Try this in:
This prompt has blanks to fill in, like [FILE]. We'll copy it instead of filling it in, so you can edit it before you send.
The source used Claude (Excel), which can take actions in a browser or app. These are general chat tools, so results may differ, and a normal chat may only give you a plan.
Opens in a new tab. We don't see what you send.
These links are only a shortcut. We're not affiliated with OpenAI, Google or xAI, and your results may not match the source.
The prompt they actually used (our rebuild)
The source didn't share its exact prompt, so we wrote this from what it describes.
Bin the variables, calculate Information Value, do variable clustering, select features, build a model in Excel and generate a ROC curve.
Optional stronger version
Our rewrite with extra checks. The poster didn't test it.
Run this task now, exactly as written below. Everything you need is here, so start straight away and only stop if something required is missing. Goal: Build an end-to-end credit risk scorecard inside Excel that follows the methodology exactly, step by step. Inputs (fill in or confirm): - Dataset ([FILE]) and target (good/bad) column - Required steps and whether everything must be Excel-native - Variables to include or exclude - Excel version Rules: - Before building, write the methodology as a numbered checklist (binning, WoE, IV, variable selection, logistic regression, scaling to points) and get my OK. - Do each step on its own sheet, in order, with formulas visible; no external tools or macros unless I allow them. - Tick each checklist item only when its sheet exists and matches the brief. - Document each variable's bins, WoE, IV and points with a reason for inclusion. - Test the scorecard on 10 sample rows and show their scores. What went wrong before, and how to avoid it: - Last time it looked impressive but broke the brief: lock the methodology as a checklist first and audit against it at the end. If you get stuck: If a step can't be done natively in Excel, stop and tell me instead of substituting another method. If a login, CAPTCHA or two-step code appears, stop and ask me to handle it; never guess or store passwords. Verify before you finish: At the end, compare the finished workbook against the checklist line by line and list any deviation. Output: Workbook with one sheet per step, the ticked checklist and a deviations list (ideally empty). Done when: every step of the brief is present, in order, and nothing was swapped for a shortcut.
Source
https://x.com/akshen121/status/2031029965605699699
Related cases
- Analyze seven years of credit card statements with Claude CoworkRecreated promptEvidence: Screenshots or logs
From seven years of monthly credit card statement PDFs in zip files, Claude Cowork returned an Excel file with 5,726 extracted, categorized transactions plus analysis; issues were vendor transaction IDs and some miscategorized vendors it wasn't told how to handle.
Poster says:5,726 transactions categorized
Claude Cowork · Data · Source date:
- Build a cohort retention table in Excel with Copilot Agent ModeRecreated promptEvidence: Article or interview
Given a long table of customers by period, Excel's Copilot Agent Mode built a color-coded cohort retention table that DataCamp's reviewer found accurate to the data; it also added a useless chart he had to delete.
Microsoft Copilot Agent Mode (Excel) · Data · Source date:
The poster says he asked Grok to graph Iowa traffic deaths per month for the last five years, then California, to check whether a roadside death counter looked low; he shared the charts.
Grok · Data · Source date: