← Back to all free guides

Never Trust Claude’s Output
Without Doing This

One agent builds. A second agent checks its work.

Claude can build you a budget forecast, a financial model, a long report or a data analysis. It can also make mistakes, like wrong formulas, numbers that don’t add up and assumptions it never told you about.

Most people trust the output and send it. That’s how errors end up in front of your boss or your client.

The Fix

After Claude finishes something important, open a second agent whose only job is to audit what the first one built. It doesn’t create anything. It recalculates the numbers, traces the formulas and flags anything off. One agent builds, one agent checks, and you end up with something you can trust.

How To Do It

01

Have Claude build the thing

Use a regular chat or a Cowork task to make whatever you need, like a budget, report, spreadsheet or proposal. Let it finish completely. (On Pro, Max and Ultra, Claude is combining chat and Cowork into one experience, so any chat can do it.)

02

Start a fresh chat or task

Don’t ask the same conversation to grade its own work, because it already believes its answer. Open a new chat and attach the file, or start a new task pointed at the folder where Claude saved it.

03

Paste the audit prompt

It checks the file against the original request and saves a list of every issue it found, without touching the original.

The Audit Prompt
You are a senior auditor. Review the file I'm giving you and find every error, inconsistency or problem. You are NOT creating anything new and you are NOT rewriting it. You are only checking what was already built. The file: [ATTACH IT, or name the file in my connected folder] What it's for: [e.g. "Q1 budget forecast going to my boss Friday"] The original request or source data it should match: [PASTE THE BRIEF, OR ATTACH THE SOURCE FILES] Check for: 1. Numbers that don't add up (totals, subtotals, percentages, dates) 2. Formulas that point to the wrong cells or give wrong results (recalculate the important ones yourself) 3. Assumptions that were made but never stated 4. Data that's wrong, missing or doesn't match the source 5. Places where the file contradicts itself 6. Anything that doesn't match the original request 7. Formatting problems that would look unprofessional For every issue, give me: - What the error is - Exactly where it is (sheet and cell, page, section or paragraph) - The correct value or fix - Severity: HIGH (wrong answer or would embarrass me), MEDIUM (should fix), LOW (polish) Rules: - Sort the list HIGH first. - If you can't verify something, say "couldn't verify" and why. Don't guess. - If you find nothing wrong in an area, say so. Don't make up problems. - End with a one-line verdict: "Safe to send", "Send after fixes" or "Do not send". Save the audit as a separate file next to the original, named [ORIGINAL NAME]-audit. Do not edit the original.

04

Fix, then check again

Take the HIGH issues back to the first chat to fix. For anything that really matters, run the audit one more time on the fixed version.

Audits For Everything Else

The same move works for anything Claude builds. Pick the one that fits and paste it into the fresh chat.

01

Budget or financial model

Financial Auditor
Act as a financial auditor reviewing [ATTACH THE MODEL OR BUDGET]. Do not change the file. 1. Recalculate every total and subtotal and list any that don't match. 2. Trace each key formula and flag ones pointing at the wrong cells, hardcoded numbers inside formulas, and ranges that miss rows. 3. List every assumption (growth rates, prices, headcount, dates) and mark which ones are stated in the file and which are hidden. 4. Stress test it: what happens to the bottom line if the two biggest assumptions are 20% worse? Output a table with columns Issue / Location / Correct value / Severity (HIGH, MEDIUM, LOW), then a one-line verdict.

02

Long report or proposal

Tough Editor
Act as a tough editor reviewing [ATTACH THE REPORT OR PROPOSAL]. Do not rewrite it. 1. List every factual claim, number and date. Mark each as supported (point to the source in the file or the material I attached), unsupported or contradicted. 2. Find places where one section contradicts another. 3. Flag any claim a skeptical reader would push back on, and say what evidence would fix it. 4. Check that it actually answers this brief: [PASTE THE BRIEF]. Output a numbered list sorted by severity with the exact location of each issue, then a one-line verdict.

03

Client-facing email or doc

Client Proofreader
Act as a careful proofreader for a client-facing [EMAIL / DOCUMENT]. Here's what the client asked for: [PASTE THEIR REQUEST]. Here's the draft: [PASTE OR ATTACH]. Check: 1. Every name, date, price and deliverable against what the client asked for 2. Anything that could be misread, sound defensive or promise more than we can deliver 3. Tone: does it sound like a confident, warm professional? 4. Typos and grammar List each issue with the exact line, why it matters and a suggested fix. Don't rewrite the whole thing. End with "Ready to send" or "Fix first".

04

Data analysis

Data Reviewer
Act as a data reviewer checking the analysis in [ATTACH THE ANALYSIS AND THE RAW DATA]. 1. Re-run the key numbers from the raw data and list any that don't match. 2. Check the method: right sample, right time period, missing or duplicated rows, outliers that skew the result. 3. For each conclusion, say whether the data actually supports it, partly supports it or doesn't. 4. Name one other explanation for the main finding that the analysis didn't rule out. Output a table: Finding / Holds up? / Evidence / Fix. Then a one-line verdict. Don't change the original files.

My Rule

If you wouldn’t send it without double-checking it yourself, have a second agent check it first. It takes 30 seconds to launch and catches things you’d miss. You still make the final call.