Claude built me a perfect file. Then I couldn’t save it anywhere
Managing a court-supervised family ledger, stress-testing Claude and ChatGPT
head-to-head to find out which one could actually update the file, not just draft it.
ROLE
Court-Appointed Administrator
PLATFORM
Claude & ChatGPT, free plan
SCOPE
ersonal, judge-reviewed ledger
TEAM
Solo
DURATION
1 afternoon
THE ACTUALS
THE PROBLEM
As a court-appointed administrator
A Claude free plan
I’m the court-appointed administrator for my mother’s care, what Italy calls an amministratore di sostegno. Part of the job is a ledger. Every euro in, every euro out, split between official and unofficial income, receipts filed against it. A judge reads this. It has to be right.
I opened Claude, free plan, and asked it to update the tracker. It understood the request immediately. It built me a clean, correctly structured file, genuinely better formatted than what I’d started with. Then I tried to get that file back into the actual Google Sheet my family uses.
Two weeks ago I needed to update that ledger, which lives in a shared Google Sheet. I assumed the newest AI model would just handle it, read the file, make the edit, save it back to Drive. That assumption cost me an afternoon.
Not through Claude directly. Not through the Chrome extension. On the free plan, there was no path from “here’s your beautifully rebuilt file” to “and now it’s live in your Drive.” I could download it. I could admire it. I could not ship it to where it needed to live.
I switched to ChatGPT, also free plan, same task. It updated the doc in Google Drive directly, no detour. Not smarter. Not better reasoning. Just wired into the one integration I actually needed that day.
IMPACT
Two tools, one real test
On a free plan, with a judge-reviewed ledger on the line, here’s what actually shipped, not what looked best in the first draft.
TOOLS TESTED
2
Claude and ChatGPT, both free plan
COULD WRITE BACK TO DRIVE
1 of 2
Only one tool finished the
actual task
TIME LOST
1 afternoon
Chasing an integration that wasn’t there
TOOLS I SPLIT BY TASK, NOW
2
One for drafting, one for the
handoff
THE TEST
Read vs. write, on the free plan
Two weeks ago I needed to update that ledger, which lives in a shared Google Sheet. I assumed the newest AI model would just handle it, read the file, make the edit, save it back to Drive. That assumption cost me an afternoon.
THE STAKES
Why this matters more than it sounds like it should
I wasn’t testing chatbots. I was doing a task with a judge on the other end of it. And I’d made the mistake a lot of us make when we pick AI tools for real work: I judged the tool by the quality of what it produced, not by whether it could actually get that output to where it needed to go.
Those are two different tests. A model can reason beautifully and still leave you stranded at the last step — the save, the sync, the handoff back into the system you actually live in. That gap is invisible until you hit it mid-task, usually at the worst possible moment.
This is the same mistake I see PMMs make with AI tools for real production work, not demos: we test the impressive part (drafting, summarizing, restructuring) and skip the boring part - does it write back to the tool I already use (Sheets, Notion, the CMS, the CRM) on the plan I’m actually paying for, or not paying for.
THE METHOD
What I actually tested
Before I build a workflow around a tool now, I test the plumbing before I trust the polish:
-
Can it read the file, or only see what I paste in?
-
Can it write back to the same place, or only hand me a download?
-
Does that write-back survive on the free plan, or does it quietly require an upgrade I haven’t checked?
None of that shows up in a first demo. It only shows up when you’re mid-task and the deadline is real.
STRATEGIC FOCUS
Better tested now than trusted blindly later
Splitting AI tools by what each can actually finish isn’t a permanent verdict on either one - it’s just where the evidence points today.
NOW
focused effort
Draft and restructure in Claude, the thinking is still stronger there for me
Hand off anything that must land in Drive to ChatGPT, free plan
Check read and write-back before trusting a new tool with a real file
LATER
parked, not abandoned
Revisit Claude for this workflow once free-plan Drive write-back ships
Check whether a paid tier actually closes the gap
Turn this into a standing checklist before adopting any new AI tool for client work
THE REAL TEST
Find the one tool that can take a file from draft to filed - not just the one that looks best doing the first half.
If you’ve hit a version of this gap - a tool that nails the output and fumbles the handoff - I’d like to hear which one and where. That’s the kind of detail that never makes it into a product page.
Reply or comment with the AI tool gap that cost you real time.