SpaceX 5-Step Process for Software Teams

Before your team speeds up or automates a process, ask who needs it, and delete it if nobody does. SpaceX builds rockets with five steps in that order, and they work for software teams too.

  • Put a person's name on every requirement, and question it even when a smart person asked for it.
  • Delete what nobody can explain, and keep a decision log so you can add it back in minutes.
  • Take extra care with security, audit and backup checks, and never delete what a regulator requires.
  • Don't simplify or speed up a process that should not exist.
  • Automate last, AI agents included, because an agent repeats waste faster than a person.

Every engineering team has a loft that is almost full. Nobody plans it. Each incident or audit adds a box: a new approval, a new ticket field, a new CI job, a new weekly call. Each box made sense to someone once. Years later the loft is full, and nobody remembers what sits in half the boxes.

Elon Musk has a method for the loft. He walked through it on a tour of Starbase, SpaceX's rocket site in Texas. Tim Dodd filmed the tour in summer 2021 for part one of Starbase Tour with Elon Musk on the Everyday Astronaut channel. The five steps start about 13 minutes in. Walter Isaacson's biography Elon Musk (2023) calls it "the algorithm". Musk built it for rockets and car factories, but it fits my engineering teams well.

The order matters more than the steps. Musk insists you run them in this sequence:

  1. Make the requirements less dumb.
  2. Delete the part or process.
  3. Simplify or optimise.
  4. Accelerate cycle time.
  5. Automate.

Make the requirements less dumb

Musk says every requirement is wrong to some degree, no matter who gave it to you. He warns that requirements from smart people carry the most risk, because you are less likely to challenge them. Smart people pack the boxes you never open.

In a software team, this step starts with a name. In Isaacson's book, Musk asks that each requirement come from a person, not a department. "Security needs it" or "Product wants it" is a box with no label. You can't ask a department why it needs something; you can ask a person. That person may think differently now, or may no longer work there.

Look for requirements like these:

  • the release approval one director asked for after an outage years ago
  • the ticket field nobody reads
  • the CI job that checks a service the team retired
  • the status report that goes to a channel nobody opens

Put a name on each one. If nobody claims a box, you have your first candidate for step two.

Delete the part or process

Musk says people add things in case they need them, and you can make that case for almost anything. His rule is that if you don't add back at least 10% of what you delete, you did not delete enough.

For a software team, the boxes are tickets, approvals, standing calls, CI jobs, dashboards and dead features. Code has a running cost long after the first version ships. Someone has to read, update, patch and test it, often after its author moves on. Process works the same way: a weekly call costs its length times its attendees.

About every two quarters, I clear my calendar and start from scratch. A call goes back in only if I can say what it is for. The same goes for the team. When someone in a retrospective says they don't see the value of a call, I ask questions first, then recommend we delete it. If we feel the pain, we can always bring it back.

There is a fair objection to all these deletes: Chesterton's Fence. G.K. Chesterton described a fence across a road. One reformer sees no use for it and wants it gone. A wiser one says to find out why it went up first, and only then take it down.

Step one already helps here: it puts a name on each box, and that person is usually the fastest way to the reason. If the reason no longer holds, delete it. If a careful look turns up no reason at all, I part ways with Chesterton. A box nobody can explain still costs something every week. I delete it anyway, then keep a closer eye on it.

Make each delete cheap to reverse. Write down what you removed, why and who agreed. My teams keep a decision log for this: we write down today's decisions for future us, and for whoever looks after the system next. It works like a note pinned to the fence, so the next person finds the reason and does not have to dig. Then the add-back takes minutes, and the 10% rule no longer feels like a threat.

The 10% rule assumes a wrong delete is cheap and easy to see. In a factory, a part left out often shows up in the next test. In software, some wrong deletes show no sign until they hurt: a security review, an access check, an audit log, a backup restore drill. You find out after the breach or the data loss. One add-back in ten can mean one incident you can't undo.

Regulated work adds a second limit. Some requirements carry a regulator's name, not a person's, and you can't delete them. Teams that ignore this delete the smoke alarm with the boxes.

Don't optimise what should not exist

Musk calls this possibly the most common error of a smart engineer: to optimise something that should not exist. He blames school for part of it. You learn to answer the question in front of you, not to question it.

So a good engineer gets a slow test suite and makes it twice as fast. Nobody asks whether half the tests check anything that matters. A team rewrites a report generator for a report nobody reads. Another team shaves minutes off a deploy that still waits a day for an approval with no owner. Each of these is a better shelf for a box that belongs in the bin.

A business partner recently wanted to automate a process to fix it. I recommended we run it by hand a few times first. Then we would meet again, look at what we learned and decide what to do next. You can't optimise what you don't understand.

Speed comes fourth

Only now do you go faster. Musk put it in one line:

"If you're digging your grave, don't dig it faster."

A fast pipeline that ships a process nobody needs wastes more per hour.

At Salesroom, we spent a few hours, maybe a day, on a plan before we tested our assumptions with users in production. Our goal was to go from problem to first solution in a week or less. I wrote about it in How many weeks do you spend planning for a feature.

Automate last, especially with AI

Automation is the final step. Bill Gates gave the reason in two rules in The Road Ahead (1995), and I come back to them often:

"The first rule of any technology used in a business is that automation applied to an efficient operation will magnify the efficiency. The second is that automation applied to an inefficient operation will magnify the inefficiency."

CI/CD, bots and AI coding agents are all automation. An agent that fills in a ticket template and updates three dashboards repeats the waste faster and never complains. Nobody pushes back on a task a machine does, so nobody thinks to open the box.

AI also makes code cheap to write. It does not make code cheap to keep. Every line an agent writes goes into the loft and carries the same running cost as a line a person writes.

In the same tour, Musk says he made this mistake himself, more than once. On the Tesla Model 3 line, he ran the five steps in reverse: automate first, delete last. Before you point an agent at a process, take that process through steps one to four.

Start with one box

Pick one standing call, approval or CI job on your team this week. Find the name behind it. Ask that person whether the team still needs it, and delete it if it does not.

Which box in your team's loft has no name on it, and what breaks if you delete it on Monday?