I built myself a skill that fills in forms — including Hebrew government forms, with the little digit boxes for an ID number and the yes/no checkboxes. One prompt built it. Two things came out of that: every PDF connector I had previously installed turned out to be unnecessary, and the part that makes the file actually useful is not the code.
The conclusion up front: one prompt gets you a skill that works. What turns it into a skill you can trust is precisely the part that isn’t code — the rules about what it must not fill in, and the precision that has accumulated in it. One line in there even says of itself that it was written only after I corrected the model.
Two steps, and most of the first one isn’t AI
The first step isn’t technical at all. Do you have a folder on your phone where you keep photos of your certificates and official documents? If not, that’s the step — and you can hand that to Claude too: go through the folder, delete anything that has expired, catalogue each scan under a clear name, and upload the lot to Google Drive. Mine came to about 26 scans — identity documents, passports, licences and certificates, and a few family papers.
The second step is a single prompt to Claude Cowork with an active Google Drive connector: build a skill that fills forms on my behalf, ask me for all the general details, produce an artifact I can scribble a signature into and save it, and look in the Drive folder to work out which documents are available to you there. For the PDF side I gave it a recommendation — install PyMuPDF — and said explicitly that if it found a better solution, it should take it.
That’s the whole recipe. The interesting part is what came out the other end.
The connectors were unnecessary
I had previously installed several connectors for manipulating and filling PDFs. Not one of them was needed. Claude just runs pip install pymupdf pillow inside its own environment and works on the file directly — reading coordinates, stamping text, drawing crosses, rendering images to check itself.
One detail brought this home. The Hebrew font the skill reaches for sits at a Linux path, which means that part isn’t running on my Mac. And there is a line in the file explaining why that’s the better direction anyway: some PDF MCP servers’ page-rendering tools fail on certain files, and PyMuPDF’s get_pixmap is the fallback that works.
The one connector genuinely required is Google Drive, and even that isn’t for the PDF — it’s for the scans. When a form asks for a passport number, a licence expiry date or a bank branch number, the skill doesn’t guess: it searches the folder, opens the right scan and reads the value off the image. The rule written there in as many words is that it must never reconstruct a value from memory and never guess a check digit. If the scan is unreadable or the document isn’t there, it asks me.
The part nobody writes in the first prompt
Israeli government forms are usually flat PDFs. No form fields, no AcroForm — just ink on virtual paper, where everything you fill in is stamped at absolute coordinates. So the first rule in the skill is to never guess a coordinate but to extract it from the file: read the label text along with its bounding box, and find the table rules and checkboxes through the drawing objects — because many forms draw their rules as very thin filled rectangles rather than as lines. A checkbox is a near-square rectangle roughly 5 to 10 points a side.
Three problems accumulated in there. One of them, according to the file itself, after a correction of mine:
Which box belongs to which label. On Hebrew forms the box usually sits to the left of its label — the word first, its box to the left of it — but the skill forbids assuming it, and requires rendering a zoomed crop of the row and looking. A mark in the wrong column on a legal declaration is not a cosmetic bug.
Direction inside a single field. Hebrew text needs bidi handling, so the skill stamps through insert_htmlbox with an embedded font rather than insert_text. But numbers, ID digits and email addresses are left-to-right even inside a Hebrew form — and the line explaining how an email address sits around a pre-printed @ was written, according to the file itself, after I’d corrected it.
A mark that doesn’t depend on the font. Instead of typing a tick or an X character, the skill draws two crossing lines inside the box. A character may not be covered by the font and comes out as a tofu square; two lines always come out as a cross.
The signature is stored not as an image but as pen-stroke geometry — a sequence of points in a 640 × 200 space, one continuous stroke, pen width 2.4. That is what lets it re-render cleanly at any size: a quadratic curve sampled through the midpoints, drawn at eight times the resolution and then scaled down so the edges come out smooth. On an A4 form it sits about 100 points wide, centred on the signature rule with its baseline resting on it.
The last step in the skill is to render the page, render a zoomed crop of the regions that were filled, and actually look at the images before handing the file over. Coordinate work breaks quietly.
What it refuses to fill
This is the part I’m most glad is in there.
Identity fields freely. Substantive questions, no. Name, ID number, address, phone, occupation, date: these are facts about me, and entering them is transcription. A question about health, money, legal history, consent or preference is a declaration I am making, sometimes under penalty of law. The skill is required to ask me about each of those separately, even when the answer looks obvious.
Third-party blocks stay empty. “To be completed by the physician”, blocks for the authority’s own internal use, an employer’s or a witness’s section — filling those in would be forging someone else’s attestation. The skill points them out and tells me at the end what is still open and who has to complete it.
Two smaller rules in the same spirit: the original file is never overwritten, only saved alongside as a new copy; and the signature is stamped only when I have explicitly asked for the document to be signed. Signing is my act, not a default.
So no, you won’t stop filling in forms
The headline on the original post was a joke and it’s worth saying so out loud. What disappears is the transcription: hunting for the passport number, matching a digit to its box, getting a signature into the file. What remains is precisely the part that requires you — the declarations. Which is fine, because it’s also the part where a mistake costs the most.
The general lesson, and it holds in organizations as much as in my own paperwork: the first prompt buys you a working skill. What turns it into something you can let near anything real is two things the first prompt doesn’t produce on its own — the precision that goes in after it gets something wrong, and the list of what it is not allowed to do. Both are worth writing into the file rather than hoping the model works them out for itself.
If you’re weighing up where in your business there’s transcription work that could move into a skill — and where the line sits that it must not cross — this is how I work.