Running a CNC shop made me a better operator

The part of my work that surprises people is the furniture.

Zeder Cabinet Studio designs, manufactures and installs custom kitchens, closets and dressing rooms in Culiacán, Sinaloa. There is a showroom, a factory with CNC machinery, and a crew that installs. The process runs in five steps: consultation, space analysis, custom design, CNC fabrication, then installation and delivery. Nothing about it is software.

It is also the single best thing that has happened to how I think about running anything, including the software.

A missed cut is not a hotfix

In software I had absorbed a set of assumptions so thoroughly that I could not see them. The main one: mistakes are cheap and reversible. Ship it, watch it, fix it. Revert if it is bad. That assumption is so deeply true of web software that it stops feeling like an assumption and starts feeling like the nature of work.

A sheet of melamine cut to the wrong dimension is a different category of event. The material is consumed. The money is spent. The machine time is spent. If the client's kitchen was scheduled for Thursday, Thursday is now a conversation rather than an installation.

The first few times this happened I looked for the process failure, the way you would after an outage. Usually there was not one. The design had been approved with an error in it, and every step downstream had faithfully executed the error at increasing cost.

That is when I understood what design review is actually for. In software, review catches bugs before users see them. In manufacturing, review catches them before material is consumed — and the gap between those two is entirely a question of how expensive your undo button is.

Material has a lead time

The second lesson took longer to accept.

You cannot compress a lead time by wanting the outcome more. Hardware ordered on Monday arrives when it arrives. A finish has a cure time. A crew that is installing on the other side of the city on Wednesday is not also installing here on Wednesday.

Software has lead times too. They are just easier to hide, because the constraint is usually a person's attention rather than a truck, and attention can be borrowed against. You can pull an evening forward. You can start the next thing before the last one is finished. The bill arrives later and in a different currency, so the connection is easy to miss.

Running a shop where the lead times are physical made me stop pretending the software ones were not real. My planning got less optimistic and more accurate, which felt like a downgrade for about a month and then stopped feeling like anything.

Done means installed

The definition of done in the shop is not "fabricated." It is not "delivered." It is installed, aligned, doors adjusted, and the client has seen it and is satisfied.

Everything before that is inventory. Inventory is work you have paid for and not yet been paid for. A shop full of finished cabinets waiting on an installation slot looks like productivity and is actually capital sitting still.

Software has exactly the same trap under different names. Merged is not shipped. Shipped is not adopted. A feature behind a flag that nobody has turned on is a finished cabinet in a warehouse. I had known that as a slogan for years. Standing in a room with the physical version of it is what made me act on it.

What transferred

Concretely, a few things changed in how I run the software and video work:

• Batch by setup cost, not by task. Changing a tool on a CNC machine costs time, so you sequence cuts to minimise changes. Context switching costs the same way. I now group recording days, editing days and development days rather than interleaving them, and the throughput difference is not subtle.
• Limit work in progress. A shop floor teaches this physically — you can see the pile. Five projects at 40 percent finished occupy space, tie up material, and produce nothing.
• Verify before you consume. Measure the room again before cutting. Confirm the assumption before building on it. The cost of the check is always smaller than the cost of the rework.
• Schedule against the constraint. In the shop it is machine time and crew availability. In the studio it is my own recording hours. Everything else is arranged around the scarce resource, not the other way round.

What does not transfer

I want to be careful here, because the tidy version of this essay would say that manufacturing discipline should be applied to software, and that is only half right.

Software's cheap reversibility is a genuine advantage, not a moral failing. Being able to ship something imperfect and correct it in an hour is a real capability, and importing manufacturing caution wholesale would throw it away. The teams I have watched do this end up with a review process that costs more than the mistakes it prevents.

The useful transfer is narrower. It is knowing which decisions in software behave like a cut — the ones you cannot take back cheaply. A database migration on live data. A public API shape. A pricing model your customers have already signed. A permissions model, once real accounts depend on it.

Those deserve the shop's caution. Most other things deserve the web's speed.

Knowing the difference is the whole skill, and I did not learn it from software. I learned it from a room where the mistakes are made of wood.