My Favorite Codex Skill Is a Dumb Little File Clerk
Bruce Hart
The best AI automation I have built lately is also one of the dumbest: a Codex skill that cleans up my Google Drive inbox.
Not dumb as in useless. Dumb as in deeply unglamorous.
It does not launch a product. It does not write a novel. It does not replace a department. It looks at a pile of scanned PDFs, bills, pharmacy receipts, house paperwork, kid art, and random documents I wanted to save, then figures out what each file is, what it should be named, and where it should live.
That is it.
And somehow that tiny thing has changed my relationship with a chore I used to avoid for months.
The skill is here if you want to see the shape of it: organize-gdrive-inbox.
Personal automation gets interesting when it absorbs judgment
The old version of this workflow was simple in theory and awful in practice.
I had one folder in Google Drive where everything landed. If I scanned a paper, saved a bill, exported a receipt, or wanted to keep an email attachment, I put it there. That part was easy. The hard part was later, when the folder had become a small administrative landfill.
Every file required a little decision.
What is this? What date should it use? Is this a bill, receipt, statement, notice, medical thing, house thing, kid thing, pet thing, tax thing, or some weird one-off? Should the filename include the amount? Is this HSA eligible? Did I already file another page from the same document? Where did I put this kind of thing last time?
None of those decisions is hard. That is what made the chore so annoying.
It was not difficult enough to feel meaningful, but it was too judgment-heavy to automate with a normal script. A rule-based script can move *.pdf somewhere. It cannot reliably look at a scanned pharmacy receipt, a mortgage document, a utility statement, and a child's drawing and make the small contextual call a person would make.
That is the sweet spot for a personal Codex skill.
Not replacing a job. Replacing a recurring micro-burden.
The inbox became an API for future me
The trick was not making Google Drive fancy. It was making the messy human input boring.
I scan papers, save emails, and dump files into one folder: my Drive inbox. I do not try to organize at capture time. Capture time is when I have the least patience and the highest chance of creating three half-baked folder systems.
Then every week or two, I run a simple Codex prompt that names the skill. Codex enumerates the folder, reads the files, uses text extraction or OCR when it needs to, and applies the rules in the skill.
The output is exactly the kind of thing I want from an agent:
- what moved
- what each file was renamed to
- where it went
- what seemed ambiguous
- what HSA-eligible expenses were found
- where the logs were saved
The skill has strong opinions because I have strong preferences hiding under years of Drive clutter. Dates go first. Vendor names get normalized. Amounts go into filenames when they matter. Health receipts go under Health by year. House maintenance has its own destination. Statements are grouped by vendor. Nothing gets deleted unless I explicitly ask for that.
That last part matters. I want automation that is willing to do work, but I do not want it casually destroying records.
So the workflow is not "AI, please manage my life."
It is more like: "Here is the inbox contract. Here is the filing policy. Apply it carefully and show your work."
The HSA table is the part that made it stick
The sneaky valuable part is not the renaming. It is the HSA output.
A lot of medical spending is easy to miss in the moment: pharmacy receipts, vision expenses, dental visits, small reimbursements, odds and ends. Individually they are not a big deal. Over a year, they can become real money if you are tracking HSA claims.
The skill does not just file those receipts. It emits a tab-separated table in the format I need for my tracking sheet: dates, patient, provider, purpose, amount, and whether there is a receipt.
That is the difference between "the receipt is somewhere in Drive" and "I have a usable row in the spreadsheet."
The boring data exhaust becomes structured.
This is where I think a lot of personal AI tooling will land. The win is not that the model can read a receipt. The win is that it can connect the receipt to a personal workflow with just enough local context to be useful.
Most consumer software cannot know my folder taxonomy, my reimbursement spreadsheet, my naming preferences, and my tolerance for ambiguity. Codex can, because I can write those preferences down as a skill and let it operate in my environment.
Skills are better than prompts when the task repeats
I could prompt this from scratch every time.
"Please look through this folder, OCR the PDFs, classify them, rename them like this, move them into these destinations, produce a log, and extract HSA receipts in this exact TSV format."
That would work once. It would also be tedious, inconsistent, and easy to get slightly wrong.
Putting the workflow in a skill changes the unit of reuse. The prompt becomes small because the policy lives somewhere durable.
That is the important mental model for me: a skill is not just a saved prompt. It is a little operating procedure for an agent.
It says when to use it. It says where the files are. It encodes the folder rules. It defines naming conventions. It sets safety boundaries. It specifies the final reporting format. It turns a vague personal preference into something executable enough for Codex to follow.
There is still judgment involved. OCR can be messy. Dates can be ambiguous. Receipts can be unclear. The model can make a wrong call.
But the workflow is now reviewable. I get a change log. I can inspect the moves. If a category is wrong, I can update the skill so the next run is better.
That feedback loop is more useful than chasing one perfect prompt.
The future might be full of tiny clerks
This is not a revolutionary app idea. Honestly, it is probably too personal to be a good app.
That is why I like it.
A lot of the best uses of agents may be too specific, too local, and too boring to become venture-backed software. They are weird little household and business workflows that sit below the threshold where you would buy SaaS, but above the threshold where you want to keep doing them manually.
Organize this inbox.
Summarize these school emails into the family calendar.
Turn these receipts into reimbursement rows.
Compare this bill against last month.
Rename these photos using the event date and people in them.
None of that needs a new platform. It needs a model with tools, access, memory in the form of written policy, and a human who still owns the result.
The funny thing is that this kind of automation feels less like building software and more like gradually hiring tiny clerks for the parts of life that leak attention.
That sounds silly. It is silly.
It also saves me time every couple of weeks and keeps a bunch of personal admin from piling up into a Saturday-sized problem.
I am trying to pay attention to that feeling. The valuable AI stuff in my life is often not the impressive demo. It is the thing that quietly removes a repeated decision from the week.
This file organizer is small, specific, and a little ridiculous.
Which is probably why it works.