Notes on Claude from the Monastery
Process
Claude has me thinking about process. If I can describe a process, I can execute it somewhat mechanically: 1) Plan steps 2) Execute next step 3) Test the step 3) Reflect and improve 4) Loop 2-4 until all the steps are complete
Process thinking existed before Claude (Toyota has entered the chat). For each step, an engineer can reach into their bag of tools to perform the step as best they can:
- Plan the car? Deploy equations about structural loads, theories of safety systems, hopes of fuel efficiency. Consider what’s worked in the past.
- Construct the car? Hire reputable construction crews; source reliable materials; build a factory.
- Test whether the car is safe? Start the car, measure the lights. Crash a couple cars on purpose to see how they behave.
- Reflect and improve? Try variations of elements and check to see if they’re better. Instill culture of constant improvement in every factory worker and engineer.
An engineer’s bag of tools is much broader than equations or code editors. Claude lets us write code faster. So why does it feel different?
Claude’s powers are broad, especially when allowed to access the whole of someone’s computer. Claude is not a tool to automate writing code. Claude is a tool to automate process.
Claude automates the process
If you can describe a process in a series of steps and have Claude reasonably execute on each of those steps, you can automate the process. Today was the first day I felt my speed of execution bend up exponentially. I saw someone make Claude automate a task we have to do every day: generate a list of PRs to send in our group’s discord chat. They deployed the skill in under an hour and will never have to fiddle with collecting all of the PRs again.
Claude teaches me to think simply and ruthlessly. What must be true for this process to succeed? Can I remove any steps without the process breaking?
As a small example, I’d love to build a Claude Skill to know when the next subway train was arriving at my stop. Previously, I’d need to 1) search for relevant data sources and APIs, 2) understand how they work, 3) write the code that correctly uses these APIs, and 4) trigger the script whenever I’m curious. Claude handles steps 1-3 now; I can delegate the drudgery to my unsleeping assistant and trigger the skill whenever I’m curious (or automatically and constantly!)
I find myself struggling to enunciate exactly what’s so special about Claude Code, as opposed to previous iterations of engineering speedup. The slide rule sped up the engineer; the IDE sped up the developer. Perhaps it’s simply speed: you don’t need to write lines of code anymore. It’s about understanding the patterns and designs.