914,400.
That's the number of "English Metric Units" in one inch. EMU is the actual name. It manages to be wrong in two languages simultaneously. And it's the unit that the Google Slides API uses for everything: positioning, sizing, transforms.
I tell you this not because you need to know about EMUs (you don't, and I'm sorry), but because this project taught me the first genuinely useful heuristic for working with AI: point it at the weirdest API first. That's where you save the most time.
The Documentation Tax
Every API has a ramp-up cost. The hours you spend reading docs, learning conventions, figuring out which StackOverflow answers are from this decade. I call it the documentation tax, and you pay it before writing a single useful line of code.
The Google Slides API has one of the highest documentation taxes in existence. EMUs for positions. Points for fonts. Normalized 0-1 floats for colors, not 0-255 like a normal API designed by people who like their users. Three different unit systems in a single API request.
I said "build me a slide with a text box at the top." Claude produced this:
'size': {
'width': {'magnitude': 8000000, 'unit': 'EMU'},
'height': {'magnitude': 600000, 'unit': 'EMU'}
},
'transform': {
'translateX': 311700,
'translateY': 445025,
'unit': 'EMU'
}
That text box is approximately 8.75 inches wide, positioned 0.34 inches from the left. But you'd see 8000000 and 311700 and think someone was entering satellite coordinates. Standard slide dimensions: 9144000 x 5143500 EMU. Every positioning value is a number that communicates nothing to a human.
No documentation detour. No unit conversion spreadsheet. No existential crisis.
Claude had already paid the documentation tax. I walked in on day one like someone's kid who got a corner office.
The Heuristic
The real lesson is a rule of thumb I used on every project after:
If the API docs make you want to quit, that's exactly when AI pays for itself.
This applies anywhere:
- Unusual unit systems (EMUs, MIDI ticks, BLE protocol bytes)
- APIs with competing conventions in the same request
- Documentation that was last updated during the Obama administration
- Anything with "legacy" in the description and no irony in the tone
The weirder the API, the higher the documentation tax, the more time AI saves. On a well-documented, modern REST API with great examples? AI is a nice convenience. On the Google Slides API with its EMUs and transform matrices and mixed unit systems? AI is the difference between a 3-day ramp-up and a 3-minute conversation.
The Credentials Situation
My .gitignore for this project had this note:
# NOTE: credentials.json and token.json ARE committed to this
# private repo for ease of use with Claude web. This is
# intentional for this low-sensitivity project.
I committed API credentials to a git repo because it was easier. The CLAUDE.md was still the 8-line stub from Post 1. The project worked despite my setup, not because of it.
The early days were lawless.
The Takeaway
AI's first superpower isn't writing code. It's absorbing documentation. The thing you dread most, the three-day deep dive into a hostile API, is the thing AI does best, because it already did the reading. Identify the tasks where the ramp-up is steepest and point AI at those first. The rest can wait.
Also: Google, if you're reading this, 914,400 is not a reasonable number of anything per inch. Please consider therapy.