Real Agent Cases

Sort a year of bank transactions for a freelancer's tax filing with Gemini, Claude and Lumo

Posted
by Megan Ellis (XDA)
Last checked
Method:
How we verify

Evidence: Article or interviewA newsletter, journalist or interview account of what happened.

PartialThe source says it only partly worked, or had a significant problem a person had to deal with.

What went wrong: Only Claude matched the manual income total; Lumo missed about a third of the income, and Gemini was close but not exact.

Context: In a comparison test: from a head-to-head chatbot review, not a standalone agent test.

Outcome

XDA's Megan Ellis, a freelancer who files taxes three times a year, gave the same bank and credit-card exports and prompts to the free plans of Lumo, Gemini and Claude. Lumo missed a month of medication payments and about a third of the income. Gemini made no obvious errors and came within $100 of the manual income total. Claude hit its usage limit within two prompts at peak times but was the only one to match the manually filtered income total. Ellis still plans to do most of the taxes by hand.

Agent

Gemini, Claude, Lumo

Prompt Recreated prompt

The source didn't share its exact prompt, so we wrote this from what it describes.

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 or a rule below tells you to.

Goal: Sort a year of my bank and card transactions into the tables I need for my tax filing, and show your totals so I can check them.

Fill in before running:
- Bank and credit-card exports for the tax year: [UPLOAD CSV]
- Categories I need (e.g. income, medical, home office): [LIST]
- Words or payees to exclude from income (transfers, refunds): [LIST]
- My own income total, if I have one: [AMOUNT]

You're not my accountant or financial adviser. Show your working, label estimates as estimates, and tell me which points a professional or the official source should confirm.

Rules:
- Make one table per category with date, payee and amount, and a total at the bottom.
- Before giving a total, say how many transactions you included and list any you weren't sure about.
- Ask me before including or excluding anything ambiguous, like transfers between my own accounts.
- If my own total doesn't match yours, show the difference line by line rather than recalculating in a loop.
- Check facts that can change (prices, opening hours, dates, availability, rules) against the official source, and tell me anything you couldn't verify.

Output: Copyable tables per category, a checklist of tax documents, and a short list of items for me to confirm.

You'll know it worked when: every total reconciles with my bank exports and I've confirmed each uncertain item.

Try this in:

This prompt has blanks to fill in, like [UPLOAD CSV]. We'll copy it instead of filling it in, so you can edit it before you send.

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.

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 or a rule below tells you to.

Goal: Sort a year of my bank and card transactions into the tables I need for my tax filing, and show your totals so I can check them.

Fill in before running:
- Bank and credit-card exports for the tax year: [UPLOAD CSV]
- Categories I need (e.g. income, medical, home office): [LIST]
- Words or payees to exclude from income (transfers, refunds): [LIST]
- My own income total, if I have one: [AMOUNT]

You're not my accountant or financial adviser. Show your working, label estimates as estimates, and tell me which points a professional or the official source should confirm.

Rules:
- Make one table per category with date, payee and amount, and a total at the bottom.
- Before giving a total, say how many transactions you included and list any you weren't sure about.
- Ask me before including or excluding anything ambiguous, like transfers between my own accounts.
- If my own total doesn't match yours, show the difference line by line rather than recalculating in a loop.
- Check facts that can change (prices, opening hours, dates, availability, rules) against the official source, and tell me anything you couldn't verify.
- Re-check one random month by listing every transaction you counted for it.

Output: Copyable tables per category, a checklist of tax documents, and a short list of items for me to confirm.

You'll know it worked when: every total reconciles with my bank exports and I've confirmed each uncertain item.

Before you finish: re-read my goal and constraints, check your answer against each one, and list anything you assumed or couldn't confirm.

Source

https://www.xda-developers.com/using-gemini-claude-lumo-taxes/

  • How-To Geek's Dibakar Ghosh gave the same job posting and prompt to Claude, ChatGPT and Gemini, each told to ask him questions first; ChatGPT asked 30. He liked Gemini's look best, but the recruiter picked ChatGPT's: the only one-page resume, with no fabricated details, though he found it busy. He says the questionnaire step is what made it work and that AI resumes still need human oversight.

    ChatGPT, Gemini, Claude · Personal · Source date:

  • Writing a month after accepting an offer from Amazon, a new hire says one of the things that helped was using ChatGPT and Claude to mock-interview her: she fed them the job description and asked them to interview her like an Amazonian would, which surfaced questions she wouldn't have thought of. She also used AI to find gaps in her 15 prepared stories but wrote them in her own words.

    ChatGPT, Claude · Personal · Source date:

  • A filer in India used ChatGPT, with documents in a Project, as his main assistant for his income tax return. He says it was good at flagging common pitfalls and asking for supporting documents, and most valuable for checking calculations across documents. But it gave confident wrong guidance on REIT dividends and sales, misclassified some short- and long-term capital gains even with broker reports, and sometimes got totals wrong; he had to verify every important number. He'd still use it next year.

    ChatGPT · Personal · Source date: