Software Is Just Another Jig

I own a 1989 Harley-Davidson Springer Softail. I recently swapped its original carburetor for a Mikuni HSR40. The swap itself was great. What I wanted was the small stuff: float height, jets, where the choke cable goes, the order of operations.

So I did what I now do with every project in my shop. I opened a notebook in NotebookLM, fed it 44 sources about the bike and the carb, and typed one line: "mikuni hsr installation and tuning on evo engine 1340."

It gave me back a single page that looks like it was drawn by a mechanic with a fountain pen and a deadline. The Evo motor in the middle, hand-shaded. Around it, boxed panels: preparation, hardware install, bench tuning, a four-step method for dialing in the throttle, troubleshooting. A skull and crossbones next to "disconnect negative battery cable." Its title is "Waking Up the Beast."

It is a beautiful tuning guide for a carburetor I don't own.

The title says HSR42. My carb is an HSR40. Somewhere in those 44 sources was the product page for the HSR42 kit, and the tool did what you'd want a grounded tool to do: it believed its sources. I went back to look at that page. Its stored content is a Cloudflare error screen: "Sorry, you have been blocked." The notebook had read the error screen too, and filed it as a source.¹

I honestly don't remember whether I added that page or the research feature did. Neither of us checked.

Then I checked one number. The page says to set the float height at 13 mm, plus or minus 2. Mikuni's own manual for the HSR42 says 18. So the guide is wrong about the carb it describes, and it describes the wrong carb.

In my career I've been wrong in exactly this way more times than I can count: confident, well formatted, and working from the wrong page. The AI fails the way people fail. It just does it faster, and with better illustrations.

That's not a reason to put it away. It's a reason to treat it like every other tool in the shop.

Makers already know this pattern

Every shop has jigs. A jig is the guide you build so a cut you make often comes out right every time, on your saw, for your work. Nobody sells the jig you need, because nobody else has your exact problem. So you build it out of scrap in an afternoon.

For most of my life, software was too expensive to be a jig. You bought what existed, or you did the math on paper. With AI, a small tool built for one shop and one person now costs an afternoon. Robin Sloan saw this coming in 2020, writing about an app he built only for his family: "I am the programming equivalent of a home cook." Maggie Appleton later named the people doing it barefoot developers.

Makers are natural barefoot developers. We already think in fixtures, tolerances and cut lists. We just didn't have the fixture for software.

A few of my jigs

ModuBox. Alexandre Chappel, a maker I admire, designed a modular storage system called ModuBOX. I wanted plywood cases for it, in my sizes, so I built a small app on top of his plans. You tell it the cases you want. It gives you the cut list, the hardware to buy, what fits on one sheet of plywood, and layouts a table saw can actually cut, with the kerf already accounted for. I wrote to Alex about it. Very cool guy.

Plater. I restore the original nuts and bolts on my bikes with an electroless nickel kit from Caswell. The manual is excellent and full of arithmetic: the surface area of every part, how many "credits" the bath has left, when to replenish. The manual even tells you to keep a checkbook for the bath. So the app is the checkbook. It computes the areas, tracks the bath, and walks me through the thirteen steps with timers and warnings. I've run it on real builds a few times, and I no longer do chemistry on the back of an envelope.²

Moto Tracker. My collection lives in one place: restoration phases, a parts pipeline from ordered to installed, a service log that nags me when something is due. Each bike's NotebookLM infographics are pulled straight into its page, so the tuning guide sits next to the parts list.

Claude on a Kindle. Some work happens outside, next to the bike, in full sun, where a laptop screen turns into a mirror. E-ink doesn't. So I wired a 2010 Kindle's browser to drive a Claude Code session. The code is public.

And one that taught me the most by failing: a cut optimizer I grew with an AI coding framework I built. It passed every test and used eleven sheets for a job that fits on two. The tests checked that parts didn't overlap. Nobody had told them about waste. I wrote about that one separately.

Comics were always the best manuals

The infographic made me smile for another reason. It belongs to a long tradition. From 1951 the US Army published PS, The Preventive Maintenance Monthly, a pocket-sized maintenance magazine drawn as comics. Will Eisner ran the art until 1971, and the magazine lasted 73 years. In 1969 John Muir self-published How to Keep Your Volkswagen Alive, "for the Compleat Idiot," illustrated by Peter Aschwanden. It sold more than two million copies.

Both understood something most manuals forget. People fix things with dirty hands and half their attention on the part. A drawing with a big arrow beats a paragraph. Today you can get that drawing from one line of text, built from your own sources. That is new, and it's fun.

My notebook shelf now reads like a garage sale with a library card: a La Pavoni espresso machine, a 300B tube amplifier build, an offset smoker, a Tamiya buggy, an aquarium cover cut on a Shaper Origin, a pizza oven, an airbrush compressor, a pressure washer pump, an Eero Aarnio Ball Chair I wanted to authenticate, a Boler trailer, a 1997 Buell.³ Each one started the same way: put the manuals and the good forum threads in one place, and ask.

Where it breaks

Other makers are finding the same edges. A Hackaday.io write-up that tested text-to-CAD tools this summer found simple brackets came out fine, while fillets, tolerances and multi-part assemblies broke. Chatbot G-code is fine for drilling a few holes and wrong on complex toolpaths. Hackaday's comment on one vibe-coded motor-controller mod is the best summary I know: "It pays to know what you're doing."

So here is my rule. The AI drafts. The factory manual decides. Any number that touches fuel, torque, current or a spinning blade gets checked against the source before it touches the work. Makers already have this habit: measure twice, cut once. It transfers perfectly.

Start with your most annoying problem

If you run a shop and haven't tried this, don't start with anything ambitious. Start with the calculation you redo every time, or the manual you can never find when your hands are greasy.

1. Put the manuals, spec sheets and your best forum threads in one notebook.
2. Ask for the one-page version. Pin it on the wall.
3. Check one number against the manual. You'll learn how much to trust the rest.
4. When the same question keeps coming back, have an AI build you a small tool for it. A calculator, a tracker, a cut list.

That's a jig. You built it for your work, it lives in your shop, and nobody else has one.

As for the Harley: I'm leaving that guide in the notebook. It's the best reminder I have to check which carburetor I'm reading about.