Compound Engineering

Overview

Compound Engineering is the practice of systematically capturing what an AI system learns from prompts, agent runs, failures, and successful workflows so that future interactions perform better. Rather than treating each prompt or agent session as a one-off event, the idea is to turn repeated usage into a compounding feedback loop: observe what worked, codify it, and feed it back into the system.

For AI Product Managers, this matters because AI products improve fastest when teams operationalize learning instead of relying on ad hoc prompt tweaks. Compound Engineering creates a bridge between experimentation and product quality by turning prompt lessons, workflow patterns, and edge-case fixes into reusable instructions, policies, or artifacts. In practice, it supports more reliable agents, faster iteration cycles, and a clearer path from prototype behavior to production-grade systems.

Key Developments

• 2026-01-11 — Jason Shuman’s conversation with Dan Shipper highlighted Compound Engineering as a key principle for AI-native organizations: capturing prompt lessons so AI agents improve over time. The concept was framed alongside broader organizational shifts such as the move from a knowledge economy to an allocation economy and the rising value of generalists who can direct human and machine intelligence.
• 2026-02-09 — Kieran Klaassen demonstrated a Compound Engineering plugin for Claude Code CLI that made the concept concrete in software workflows. Using commands such as `workflows plan`, `workflows work`, `assess`, and `triage`, the system ran a planning–coding–assessing–codifying loop, stored learnings in local documentation folders, and updated a root `claude.md` file so those insights were injected into future prompts. The workflow also incorporated automated browser testing with Playwright and unattended execution for more continuous AI-driven development.

Relevance to AI PMs

• Turn experimentation into product memory. AI PMs often run many prompt and agent experiments that never get operationalized. Compound Engineering provides a method for capturing winning prompts, failure patterns, evaluation rubrics, and edge-case resolutions in a way that future runs can reuse.
• Improve reliability through structured feedback loops. Instead of only measuring outputs, PMs can define workflows where systems plan, execute, assess, and codify lessons. This creates a practical path to raising consistency across repeated user tasks, internal copilots, or autonomous agents.
• Build scalable AI operations. As products move from demo to production, ad hoc prompt iteration becomes hard to manage. Compound Engineering helps PMs formalize learnings into docs, system instructions, test cases, or retrieval sources, making team knowledge durable and easier to scale across use cases.

Related

• claude-code — A concrete implementation context where Compound Engineering was demonstrated, showing how CLI-based agent workflows can capture and reuse learnings over time.
• kieran-klaassen — Demonstrated a plugin-based workflow that operationalized Compound Engineering through repeatable planning, coding, assessment, and codification loops.
• dan-shipper — Helped articulate Compound Engineering as part of a broader view of how AI-native organizations should improve systems through accumulated prompt learning.
• jason-shuman — Surfaced the concept in discussion with Dan Shipper, connecting it to practical shifts in how teams organize around AI.