Compound engineering makes each unit of work improve all future work

Most engineering treats work as linear — you finish one thing, start the next. Compound engineering inverts this: each unit of work makes the next one easier. The four-step loop is Plan → Work → Review → Compound, with 80% of time in Plan and Review. The Review phase is the compounding mechanism — it’s where patterns get extracted, mistakes get documented, and Markdown skill files may replace expensive fine-tuning grow. The Compound step explicitly records learnings for future use — this is what makes it compound, not just iterative.

The 12-agent parallel review (architecture, security, performance, simplicity reviewers running simultaneously) is an implementation of Declarative beats imperative when working with agents — each reviewer has success criteria, not step-by-step instructions. The result compounds because updated CLAUDE.md files, pattern docs, and skill files mean the agent on iteration N+1 is fundamentally better than on iteration N.

This is the engineering equivalent of Persistent agent memory preserves institutional knowledge that walks out the door with employees: documentation isn’t a chore at the end — it’s the mechanism that makes compound growth possible. Compilation scales but curation compounds — two camps for knowledge graph construction makes the same point about knowledge systems: compiled wikis grow fast but hit a quality ceiling, while curated knowledge (like this graph) compounds through validated connections — the review phase IS the curation mechanism. Skip the review phase and you get linear productivity. Do it consistently and you get exponential. shadcn’s /done skill demonstrates this practically — Session capture turns ephemeral AI conversations into a compounding knowledge base turns every Claude session into searchable development memory, making the review-and-extract pattern lightweight enough to sustain daily. OpenAI’s Codex team took this to the extreme: When agents write the code, engineering moves into the harness — and documentation becomes the system of record describes how they built a million-line codebase with zero manually-written code by making documentation the primary engineering artifact — the compound loop operating at organizational scale. When applied to products rather than engineering process, Proprietary feedback loops create moats that widen with every interaction shows the same compounding dynamic: each user interaction improves the system for all future interactions. And Revealed preferences trump stated preferences — track what users do, not what they say sharpens what to compound on — track actual behavior, not survey responses. Clay calls the human-organizational equivalent Negative maintenance teammates reduce future work for everyone around them — teammates who proactively reduce future work for others are applying compound engineering to organizational friction.

Tan gives the individual-scale habit that produces the same compounding, and names it the discipline “that separates the compounders from the dabblers”: never do one-off work — at the end of every task, turn what the agent just did into a reusable markdown file, because “it sort of doesn’t matter how good the model gets if you can’t turn it into real memory.”

Tan’s version (Garry Tan — “Own Your Intelligence”, Startup School 2026, sources/garry-tan-yc-own-your-intelligence.md): most people run a single operation with an agent and then throw the context away — they close the window, and that is it. His compressed rule, which he says he uses at YC, is “If you have to ask for something twice, you failed.” “The person who captures what they learn gets smarter every single day,” while the alternative is waking up every morning with amnesia. That per-task capture is how a rented model becomes an owned asset, per Personal AGI is an asset you build, not a subscription you rent, and its accumulated output is what is at stake in Own your skill files, because if you don't, your job becomes a skill file. It inherits one caveat from Retrieval is the primitive, not the product — being worth retrieving from is the product: skillifying indiscriminately builds the garbage dump, so capture needs pruning as its other half.