I Asked My AI Assistant to Print a Transaction Form

This morning I needed to print a transaction form for my cash ledger.

The form already existed as a small HTML file in my ledger repository:

~/Personal/ledger/2026/cash/transaction-form.html

Normally this is the kind of tiny admin task that eats attention. You open the file, pick a printer, discover the wrong queue or scaling, fix one small thing, and try again. None of this is hard, but it still interrupts the morning.

So I asked my AI assistant:

open transaction-form.html from ledger repo, it’s in 2026/cash folder. open in a browser to render as PDF and print this PDF

And it did the boring work.

The flow

The agent started a temporary local web server from the cash ledger folder:

cd ~/Personal/ledger/2026/cash
python3 -m http.server 8765 --bind 127.0.0.1

Then it rendered the HTML page into a PDF with headless Chrome:

google-chrome-stable \
  --headless \
  --no-sandbox \
  --disable-gpu \
  --print-to-pdf=/tmp/transaction-form-20260514-085744.pdf \
  http://127.0.0.1:8765/transaction-form.html

Then it checked the available printer queues:

lpstat -p -d

There were two Xerox queues. One was disabled. The Bonjour queue was enabled:

[email protected]

So the agent sent the generated PDF there:

lp -d '[email protected]' \
  /tmp/transaction-form-20260514-085744.pdf

The first job failed.

This is where the agent became useful. I just said:

retry

It checked the queue, cancelled the failed job, and sent the PDF again with page fitting enabled:

lp -d '[email protected]' \
  -o fit-to-page \
  /tmp/transaction-form-20260514-085744.pdf

This time CUPS reported:

Printer is printing

Done.

The important part was not printing

Printing one PDF is not impressive.

What matters is the loop:

1. I gave the agent the intent, not step-by-step instructions.
2. It found the file in the right repository.
3. It rendered HTML to PDF instead of trying to print raw HTML.
4. It inspected the printer state.
5. It handled a failed print job.
6. It retried with the correct option.
7. It stopped the temporary server afterwards.

This is where personal agents are useful: removing the boring connective tissue between tools.

In this case that meant the shell, the browser, the printer queue, the local repo, and the remembered preference that the form should be PDF first.

Turning the fix into memory

After it worked, I asked:

extract into a skill

The agent updated my existing ledger-cash-capture skill with the reusable procedure:

• use 2026/cash/transaction-form.html
• render HTML to PDF with headless Chrome
• check printers with lpstat
• prefer the enabled Bonjour Xerox queue
• print with lp
• retry failed jobs with -o fit-to-page
• stop the temporary HTTP server afterwards

Then I added a Russian trigger too:

напечатай форму для транзакций

So next time I do not need to remember any of this. I can just ask for the thing I want.

This is why skills matter

A good agent should not only solve the task. It should get better after solving it.

The first run was exploration. The next one can follow a known path.

That is the difference between a chatbot and a useful personal automation system. The value is not only in the model. It is in the accumulated procedural memory around your actual life.

It is a tiny example, but it is a real workflow: printed paper in hand, and less printer nonsense to remember next time.