# The AI-Assisted Audit — Comeback Prompts
*(Companion to Chapter 4. Copy-paste these into Claude, ChatGPT, or any capable AI assistant. Read the privacy box first — it is not optional.)*

## ⚠️ Before you upload anything: hide your private info first

Your credit report contains your Social Security number and date of birth, and those never get uploaded to ANY AI service. The **Redaction for Regular People guide in this toolkit** shows five no-skill ways to handle it — the easiest takes two minutes and doesn't even require marking anything up (you just re-save the PDF without the personal-info pages, using the Print → "Save as PDF" trick). Short version of the rules:

1. **Turn off chat history / model training** in the AI's settings if the option exists.
2. **SSN and date of birth: never uploaded.** Full account numbers trimmed to the last 4 digits — though on AnnualCreditReport.com reports they're usually already shown that way.
3. Your name, creditor names, balances, and dates can stay — the audit needs them.
4. **Never use an online "redact my PDF" website** — that's uploading the unredacted report to a stranger to hide it from strangers. The guide's methods all happen on your own device.

If any of it feels like too much, use Prompt 1's "typed summary" variant instead — you type the accounts in, no document upload at all.

---

## Prompt 1 — The Extraction (one report → one clean table)

> I'm auditing my own credit report for errors. Attached is my [Equifax/Experian/TransUnion] report [redacted per my privacy checklist]. Extract every tradeline, collection, public record, and hard inquiry into a table with these columns: Creditor name · Account (last 4) · Type · Status (open/closed) · Reported balance · Credit limit · Date opened · Date of first delinquency (if shown) · Late payments (list each month + severity, e.g. "Mar 2024: 60") · Remarks.
>
> Rules: Do not summarize or skip "unimportant" accounts — I need every line. If a field isn't shown in the report, write "not shown" rather than guessing. If any text is unreadable, flag it as [UNREADABLE] instead of inferring. Add a "Likely company" column: where a creditor name is an abbreviation or issuing bank (CBNA, SYNCB, COMENITY, WFBNA, JPMCB, DSNB, etc.), name the company probably behind it and the kinds of cards it issues — labeled as a guess for me to confirm, never a conclusion.

*(Typed variant: "I'll type my accounts in instead of uploading. Build the same table from what I give you, then ask me for any missing fields one batch at a time.")*

## Prompt 2 — The Cross-Bureau Comparison (the tedium killer)

> Here are extraction tables from my Equifax, Experian, and TransUnion reports. Compare them and list every discrepancy: accounts appearing on some reports but not others; the same account with different balances, limits, statuses, or late-payment histories; the same debt appearing more than once (possible duplicate — original creditor plus collection agencies, or two collectors on one debt); and any personal-info differences worth checking. For each discrepancy, tell me which bureau shows what.

## Prompt 3 — The Anomaly Scan (the error checklist, automated)

> Using the table(s) above, flag every item that matches these error patterns, and for each flag explain what triggered it:
> 1. Negative items older than 7 years from date of first delinquency (assume today is [DATE] — show your date math).
> 2. Collections whose "date of first delinquency" is suspiciously recent compared to the account's opened date or my note of when I actually stopped paying (possible re-aging).
> 3. Possible duplicates of the same underlying debt.
> 4. Closed accounts reported as open, or vice versa.
> 5. Credit limits that look wrong or missing (hurts utilization).
> 6. Accounts I've marked as "don't recognize."
> 7. Any late payment on an account I've marked "always paid on time."
>
> Output as a prioritized target list: strongest disputes first (not-mine, obsolete, documented errors), and clearly label which items you CANNOT judge without my records — those are questions for me, not conclusions.

## Prompt 4 — The Letter Drafter (use with the letter pack)

> I'm disputing this item with [bureau/furnisher]: [paste the item row + what's wrong + what's true + what evidence I have]. Using plain, factual, first-person language — no legal jargon, no "under penalty of perjury" theatrics, no template-sounding phrases — draft a one-page dispute letter following this structure: [paste the relevant skeleton from the letter pack]. Keep my stated facts exactly as I gave them; do not invent dates, amounts, or claims I didn't make.

## Prompt 5 — The Response Decoder

> A credit bureau/collector sent me this response to my dispute: [paste or attach, redacted]. In plain English: (1) What are they actually saying? (2) Did they meet their legal obligations (30-day investigation, method-of-verification description, validation itemization) or dodge them? (3) Based on the Comeback playbook — follow-up letter, furnisher dispute, CFPB complaint, or done — what's my logical next move? If their letter is boilerplate that doesn't answer what I asked, say so bluntly.

---

## What AI is for — and what it's not for

**Good at:** reading 60 boring pages so you don't have to; cross-referencing three reports; date math; spotting patterns; drafting your facts into clean letters; decoding bureaucratic responses.

**Never do this:** let it decide what's *true* (only your records know if that payment was late); paste unredacted reports; ask it to "find loopholes" (you've read Chapter 3 — there aren't any, and AI will happily hallucinate some); send any letter you haven't read and verified line-by-line yourself. The AI builds the suspect list. You are the detective who confirms it. Every letter goes out with your name on it — make sure every word in it is yours and true.
