Task
I edit AI-generated drafts for clients.
Most of the drafts are articles, reports, or marketing copy. They look polished. They read smoothly. And they often contain errors that are hard to spot: wrong statistics, invented sources, missing context, dates that don’t match.
I used to track claims in my head. That worked for short pieces. It failed for anything over 500 words. I’d catch some errors and miss others. I needed a system.
So I built a spreadsheet. It’s not fancy. It’s a Google Sheet with ten columns. I fill it in as I audit a draft. By the end, I have a complete record of every claim I checked and what I found.
I’ve been using it for about two months. It’s helped. But I think it’s missing something. I’m hoping the community can tell me what.
Tool and Date
Tool: Google Sheets (free tier)
Date: September 15, 2026
Used on: Six client drafts so far
Original Input
I start with an AI-generated draft. I paste the full text into a document. Then I create a new row in the spreadsheet for every factual claim I can identify.
A factual claim is anything specific and verifiable: a number, a date, a name, a study, a statistic, a direct quote, or a prediction presented as fact.
I don’t track opinions. I don’t track generic statements. I track things that can be checked.
Claim Under Review
My spreadsheet has ten columns. Here they are.
Claim ID — a number, so I can refer back to it.
Exact claim text — the sentence or phrase from the draft.
Claim type — statistic, date, name, study, quote, prediction, or other.
Source needed? — yes or no. Most claims need a source.
Source found? — yes, no, or unresolved.
Source type — primary, secondary, news article, blog, social media, or none.
Does the source support the claim? — yes, no, partly, or can’t tell.
Finding — confirmed, corrected, partly correct, or unresolved.
Risk level — low, medium, high, based on what happens if someone trusts the claim.
Correction — the correct version of the claim, or “none needed.”
That’s it. Ten columns.
It works. But I’ve noticed gaps. For example, I don’t record which AI tool generated the draft. I don’t record the date the draft was generated. I don’t record whether the claim involves personal data or a specific jurisdiction. I don’t record whether I’ve archived the evidence.
I think I need more columns. But I don’t want to add columns that nobody uses.
Evidence
I tested the spreadsheet on six drafts. Here’s what happened.
Draft 1: 12 claims. 2 errors. Spreadsheet caught both.
Draft 2: 8 claims. 1 error. Spreadsheet caught it.
Draft 3: 25 claims. 5 errors. Spreadsheet caught four. I missed one because the source was behind a paywall and I didn’t mark it unresolved.
Draft 4: 6 claims. 0 errors. Spreadsheet was overkill, but harmless.
Draft 5: 30 claims. 8 errors. Spreadsheet caught all eight, but I spent 3 hours on it.
Draft 6: 15 claims. 3 errors. Spreadsheet caught all three.
The spreadsheet works. The missing column cost me one error in Draft 3.
Finding
Partly correct.
My spreadsheet is useful. It catches errors I would have missed. But it has gaps. I need to add fields for:
AI tool and version — so I can track whether one tool fails more often.
Draft generation date — so I can tell if the AI used outdated information.
Jurisdiction — for claims about laws, regulations, or local rules.
Privacy redaction status — for claims that involve personal data.
Evidence archive link — for the screenshot or PDF I saved.
Confidence level — how sure am I that my finding is correct?
That’s six potential new columns. I don’t want to add all of them. I want to add the ones that matter.
Risk
If I don’t add the right fields, I’ll keep missing errors.
For example, without a jurisdiction field, I might treat a local regulation claim as a general fact. Without a generation date, I might not realize the AI was working with outdated information. Without an evidence archive link, I might not be able to prove my finding if someone asks.
The risk isn’t huge for every draft. But for client work, it matters. A missed error can cost a client money or reputation. A missed pattern can cost me future work.
Lesson
Here’s what I’ve learned so far.
Ten fields is not enough. I need at least twelve. Probably fourteen.
More fields mean more time. Every extra column adds a few seconds per claim. For a 30-claim draft, that adds up.
The right fields depend on the draft. For a legal or regulatory draft, jurisdiction matters. For a marketing draft, it might not.
Unresolved is a valid finding. I used to force a yes or no. Now I’m comfortable marking a claim unresolved if I can’t verify it.
Evidence archiving matters. If I can’t show my work, my finding is just an opinion.
What I’m Asking the Community
Here’s my current template. What field would you add?
I’m specifically interested in fields that:
Help catch errors earlier
Take less than 5 seconds to fill in
Apply to most drafts, not just specific types
Make the audit easier to share or review
I’m also curious about fields you’ve tried and removed. What seemed useful but wasn’t?
And one more question: should the spreadsheet include a field for the original prompt? I’ve been leaving it out because prompts are often long and messy. But maybe that’s a mistake.
If you have a template that works, please share it. If you’ve seen mine and think it’s missing something obvious, tell me. I’d rather fix it now than after another missed error.
Comments
No comments yet — be the first to share a thought.