Start with a result from the wrong field. In 1988 the cognitive scientist John Sweller published what became Cognitive Load Theory: working memory holds only a few chunks at once — call it four, on a good day, before the coffee — and every symbol you don’t yet understand spends one of those slots. Sweller split the cost into intrinsic load (the actual difficulty of the idea) and extraneous load (the difficulty of the packaging). Teaching, in one sentence, is the art of paying down the extraneous so the intrinsic has room to land.

Now open the thing that prompted this piece. The ActiveState Python recipe index is a wall of good, specific snippets — Vasudev Ram’s class-based callbacks recipe, logged at roughly 38,000 reads over seven years, sits next to Jorj X. McKie’s PDF text extraction with fitz / MuPDF, about 35,000 reads over eight years. To a fluent developer it reads like a buffet. To an architect in their first week of code it reads like a foreign menu with no pictures: self, decorators, __init__, indentation that is load-bearing and invisible. That is pure extraneous load. None of it is the idea. All of it is the packaging.

This is the whole pedagogical defense for teaching Grasshopper before Python at PAZ Academy, and it is not a matter of taste. In Grasshopper the data flow is the wire. You see the number leave one component and arrive at the next; there is no syntax tax between the intention and the result. The designer spends their four precious slots on the geometry, not on where the colon goes. As our own PAZ concept panel on creative coding puts it, the expressive sketch and the production script are the same skill at two altitudes — and the playful altitude is where fluency is cheapest to buy. Grasshopper buys the intuition; Python spends it.

The Tool: The graduation moment — when the node graph runs out of road — is the honest reason to open a text editor, and it is exactly where PyMuPDF earns its afternoon. PyMuPDF (imported as fitz, maintained by Artifex on top of the MuPDF engine) reads the text layer, images, and geometry out of a PDF in a few lines. For a Bauleitung or a small practice, that PDF is the tender, the Ausschreibung, the spec no one will hand you as a spreadsheet. This is the first task where Grasshopper genuinely cannot help you and Python obviously can — which makes it the perfect first real script.

Setup:

python -m venv .venv
.venv\Scripts\activate        # Windows; use: source .venv/bin/activate on macOS/Linux
pip install pymupdf
python -c "import fitz; print(fitz.__doc__)"   # prints the build banner = it works

First steps:

1. Drop any tender PDF next to your script and rename it tender.pdf.
2. Create extract.py with four lines: import fitz, open the doc, loop the pages, print the text.

   import fitz
   doc = fitz.open("tender.pdf")
   for page in doc:
       print(page.get_text())
   

3. Run python extract.py. The specification you were about to read by hand scrolls past as plain, searchable text — the render smiles, but so, finally, does the person reading the fine print.

←TODAY: In 2026 an architect who can read a PDF’s text layer audits a 60-page tender in minutes, not an afternoon. →3012: The tools that survive to Zurich-3012 are the ones whose files a stranger can still open unaided. Fulcrum: Teach the visual altitude first so the text altitude lands — the graph is where you learn to think, the script is where you learn to keep.

A word from the far side of the format wars, because pedagogy is also a longevity bet. The buildings that aged badly in my time were not the ugly ones — they were the ones whose files went dark when a vendor did. PDF is an ISO standard (ISO 32000); a .py file is text a 25-year-old can still read; a proprietary node graph is a promise from a company. We teach Grasshopper first for cognitive load, but we make sure the durable output — the extracted text, the plain script, the open format — is where the value settles. If you are choosing a stack this quarter, ask the only question that matters when the vendor disappears: can a stranger still open the file?

Atelier: For a Swiss practice, the payoff is a triage skill, not a career change. A studio living with AI right now is drowning in PDFs it half-trusts — tenders, product data sheets, LOIN schedules — and it reaches for a chatbot to summarise them, which is fine until the number is wrong and unattributable. The Monday move: put a fifteen-line PyMuPDF script on one workstation and pipe every incoming tender through page.get_text() before anyone reads it, so the office searches the real text instead of guessing at a scan. One script, one afternoon, zero licences.

Hack: Sweep the whole tender for every dimension it quotes, in one pass. Once the text is out of the PDF, a single regular expression turns a document you skim into a document you audit — every mm value, deduplicated and sorted, so the odd one out is obvious.

import re, fitz
text = "".join(p.get_text() for p in fitz.open("tender.pdf"))
dims = re.findall(r"\b\d{2,5}\s?mm\b", text)   # every 2-5 digit mm value
print(sorted(set(dims), key=lambda s: int(re.match(r"\d+", s).group())))

Run it against two revisions of the same tender and the diff of those two sets is your change list — the slab that quietly went from 240 mm to 220 mm between issues, caught before the bricklayer does.

None of this contradicts the Grasshopper-first rule; it completes it. The node graph taught you that a pipeline is inputs → transform → outputs. PyMuPDF is that same shape written in text: open, transform, print. You already understood it visually before you ever met a colon — which is exactly the point. PAZ’s own Archicad-Python sessions make the same argument from the other end, that the real barrier for architects is rarely the logic and almost always the documentation gap and the syntax tax around it.

Learn-it:

• The source that started this: ActiveState Python recipes — including McKie’s PyMuPDF text extraction
• Kill the PATH problem first: Python on Windows — official docs
• The footguns your CAD script will meet: A bad day to use Python
• Dependency hygiene before you pip install: The litellm PyPI supply-chain incident
• PAZ note — the same argument from the Archicad side: Die 3 Dokumentationslücken, die alle in Archicad Python ausbremsen

So: teach the wire before the colon, and prove the payoff on the ugliest PDF in your inbox. Clone nothing you don’t need, install PyMuPDF, and make your first script the one that reads your next tender for you.