Compounding Engineering
Spend a little effort now so each repeat of a task is cheaper than the last
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 4
- Confidence
- 95%
In an AI-agent workflow where humans no longer hand-write code, most of the work becomes writing PRDs and prompts. Compounding engineering says: for every unit of work you do, invest a little extra to make the next identical unit of work easier. Instead of grinding out the same artifact repeatedly at constant cost, you build reusable prompts and automations so the marginal cost of each future task falls.
Origin
Attributed by Dan Shipper to the two-person Kora engineering team at Every (Kieran and Nateesh), who he says invented the concept; Shipper states he 'had nothing to do with this.'
Core principles
- 01Every unit of work should make the next unit of work easier
- 02In an agent-driven org your real job shifts to specifying work (PRDs/prompts), so optimize that layer
- 03Reusable prompts and slash commands are the compounding asset, stored and shared in version control
- 04Small speedups accumulate into large leverage across a tiny team
How to run it
- 1
Notice the repeated artifact
Identify the thing you produce over and over in an AI workflow — for Every it was PRDs (detailed specs) fed to Claude Code. Recognize that doing it the same way each time yields no leverage.
Pro tip The tell is 'this is just my job now' — the moment a task feels permanently manual is the moment to templatize it.
- 2
Define the platonic ideal of the output
Articulate what a great version of that artifact looks like (e.g. the ideal PRD structure) so a model has a target to hit.
- 3
Write a prompt that produces it from raw input
Build a prompt or slash command that turns your rambling thoughts into the finished artifact, so future instances take a fraction of the effort.
Pro tip Claude Code slash commands are repeated prompts you invoke on demand — a natural home for these.
- 4
Store and share the library
Keep the prompts/automations in a shared GitHub repo so the whole team compounds off each other's speedups, not just you.
In the wild
In a Claude-Code world where they weren't hand-coding, the Kora team found they spent most of their time typing PRDs. Rather than accept that as fixed-cost work, they wrote a prompt that takes rambling thoughts and turns them into a well-formed PRD, making every subsequent PRD cheaper to produce.
→ A two-person team (plus ~15 Claude Code instances) shipped Kora to 2,500 active users processing millions of emails, as one of several products at a 15-person company.
Common mistakes
Treating spec-writing as permanent manual labor
Accepting 'writing PRDs is just my job now' locks you into constant-cost work. The leverage comes from spending a little time to automate the production of the artifact itself.
Is it for you?
Best for
Small AI-first product/engineering teams working through agents who find themselves repeatedly producing the same kind of specification or prompt
Not ideal for
One-off or genuinely novel tasks where no repeated pattern exists to compound against
From the transcript
“they invented the idea of compounding engineering”
“So basically for every unit of work you should make the next unit of work easier to do”
“you end up spending a lot of time essentially typing PRDs”
“what I'm going to do is write a prompt that can take my rambling thoughts and then turn that into a PRD”
“You can do slash commands in Cloud Code, which are like repeated prompts”
From the episode
The AI-native startup: 5 products, 7-figure revenue, 100% AI-written code
Dan Shipper (co-founder/CEO of Every)