AI Underwriting Prompt Starter — 10 battle-tested prompts

If you’ve tried ChatGPT for underwriting, you know the pattern: the first prompt feels magical, the second is mid, the third is noticeably wrong, and by the fourth you’re back to doing it yourself. The problem is almost never the model. It’s the prompt.

This pack is the starter set of ten prompts I’ve actually used — and broken, and rewritten — to do real underwriting work. Not “10 ChatGPT Hacks for Bankers” copy. Working prompts, with the input they expect, the output they produce, and a note on where they still fall over.

What a prompt looks like in this pack

Every prompt is a complete one-pager:

1. Goal — one sentence. What this prompt is for and what it’s not.
2. Prompt text — the actual string you paste.
3. Input spec — what you need to paste with it (a bank statement, a job letter, a CIBIL report excerpt, etc.).
4. Annotated example — a real run with a synthetic borrower, with my notes on what the model got right and where it tried to hallucinate.
5. Tweak notes — knobs you can turn. Different model? Different loan type? Indian vs. US borrower? Here’s what to change.
6. Where it breaks — the cases where the prompt produces confidently wrong output. Because pretending these don’t exist is how people get in trouble.

The ten prompts

1. Intake normaliser — converts a messy application form into a clean structured record.
2. Borrower summary — 200 words, neutral tone, ready for a credit memo.
3. Income-document red-flag scan — doesn’t just summarise; it actively looks for the three most common doctored-payslip patterns.
4. Affordability calculator assistant — computes DTI, FOIR, and stress-tested EMI headroom from raw statements.
5. Reference-check question generator — tailored to the specific concerns this file raises.
6. Credit memo draft — structured output: risk grade, rationale, conditions, deviation request.
7. Collections email (early-stage) — firm, not harassing, regulator-friendly.
8. Customer comms rewrite — takes your internal-language approval/decline and rewrites it in plain language for the borrower.
9. Portfolio Q&A prompt — ask your portfolio natural-language questions without re-learning SQL every week.
10. Post-mortem prompt — when a loan goes bad, this prompt structures the lessons-learned write-up so the same hole doesn’t appear in the next underwriting.

Who built this, honestly

Welcome. I’m not a career underwriter — I’m a builder going deep on AI-in-lending in public. Every prompt in this pack is something I’ve stress-tested against real (anonymised) lender workflows and synthetic data. Where I don’t know, I say so. The “where it breaks” section in each prompt is not marketing polish — it’s what took me the longest to figure out.

The honest version of “why buy this”

You can write these prompts yourself. It will take you a long weekend and you will make every mistake I made. The $9 is for the weekend, and for the seven or eight non-obvious fixes I’ve baked into each prompt that you wouldn’t land on in round one.

If you want the full 50-prompt library spanning the whole lending workflow, that’s the $29 Prompt Library. If you want the library plus the scorecard template plus the fraud playbook, that’s the $149 Starter Kit.