Compound Engineering
A practice of capturing learnings from prompts and agent interactions to steadily improve system behavior over time. For PMs, it is a feedback-loop mindset for iterative AI product improvement.
Key Highlights
- Compound Engineering is the practice of capturing prompt and agent learnings so future AI outputs improve over time.
- For AI PMs, it creates a structured feedback loop connecting prompts, evaluations, documentation, and product reliability.
- Dan Shipper and Jason Shuman framed the concept as a principle for AI-native organizations.
- Kieran Klaassen demonstrated a practical implementation in Claude Code using plan, work, assess, and triage workflows.
- The concept helps teams turn one-off prompt insights into reusable system behavior and operational knowledge.
Compound Engineering
Overview
Compound Engineering is the practice of systematically capturing what works and what fails in prompts, workflows, and agent interactions so future AI outputs improve over time. Instead of treating each prompt or agent run as a one-off event, it treats every interaction as a source of reusable product knowledge—turning experiments, mistakes, and successful patterns into durable instructions, rules, and operating procedures.For AI Product Managers, this matters because AI product quality is rarely improved by model choice alone. The biggest gains often come from tightening the feedback loop between user behavior, prompt design, evaluation, and system instructions. Compound Engineering provides a practical mindset for doing that: observe outcomes, codify lessons, inject them back into the system, and steadily increase reliability, speed, and consistency across use cases.
Key Developments
- 2026-01-11: In Jason Shuman’s conversation with Dan Shipper, “compound engineering” is framed as a principle for AI-native organizations: capturing lessons from prompts and agent usage so AI systems improve over time.
- 2026-02-09: Kieran Klaassen demonstrates a Compound Engineering plugin for Claude Code CLI. The workflow uses commands such as `plan`, `work`, `assess`, and `triage` to create a planning–coding–assessing–codifying loop, storing learnings in local documentation and feeding them back into future runs.
- 2026-02-09: The demonstrated implementation appends codified learnings to Markdown files under `/docs/architecture-decisions/` and `/docs/solutions/`, while updating `claude.md` so those rules are injected into subsequent planning prompts.
- 2026-02-09: The same workflow shows how compound engineering can extend beyond prompting into automated testing and execution, including Playwright-based end-to-end browser tests and unattended Claude Code sessions.
Relevance to AI PMs
- Build repeatable improvement loops: AI PMs can formalize a cycle of prompt design, execution, evaluation, and codification so each release improves from prior runs rather than restarting from scratch.
- Turn tacit team knowledge into system behavior: When prompt lessons, edge cases, and failure patterns are documented in reusable instructions or memory files, teams reduce dependence on individual experts and make performance more consistent.
- Improve reliability with operational artifacts: PMs can treat artifacts such as eval results, test cases, architecture decisions, and solution docs as product inputs that shape future agent behavior, not just as passive documentation.
Related
- Claude Code: A practical environment where compound engineering can be implemented through iterative coding, assessment, and codification workflows.
- Kieran Klaassen: Demonstrated a concrete Compound Engineering plugin and workflow for Claude Code CLI.
- Dan Shipper: Helped popularize the idea in the context of AI-native organizations and long-term agent improvement.
- Jason Shuman: Surfaced the concept through discussion of organizational principles for working effectively with AI.
Newsletter Mentions (2)
“Kieran Klaassen demonstrates his Compound Engineering plugin for Claude Code CLI, using slash commands like workflows plan, workflows work, assess, and triage to run a planning–coding–assessing–codifying loop that captures insights in a local docs directory and iteratively improves generated code.”
#4 ▶️ How to Make Claude Code Better Every Time You Use It (Full System) | Kieran Klaassen Peter Yang Kieran Klaassen demonstrates his Compound Engineering plugin for Claude Code CLI, using slash commands like workflows plan, workflows work, assess, and triage to run a planning–coding–assessing–codifying loop that captures insights in a local docs directory and iteratively improves generated code. The compound-engineering-plugin appends codified learnings as Markdown under /docs/architecture-decisions/ and /docs/solutions/ , and updates the root claude.md so those rules are injected into every new workflows plan prompt. With Opus 4.5 and Playwright, Claude Code auto-generates end-to-end browser tests—logging into Gmail to exercise email signature and draft flows, clicking UI elements, inspecting console logs, and screen-recording a video artifact attached to the pull request. By defining alias CC="claude code --dangerously-skip-permissions" , all interactive permission prompts are suppressed, enabling fully unattended AI-driven sessions for commands like plan, work, assess, and PR creation.
“Jason Shuman’s conversation with Dan Shipper surfaces key principles for AI-native organizations: the shift from a knowledge economy to an “allocation economy” where orchestration of human and machine intelligence is paramount; the resurgence of generalists with strong taste and direction; and “compound engineering,” capturing prompt lessons to improve AI agents over time.”
Product Management Insights & Strategies Marc Baselga outlines three investor-selection filters for first-time founders: diversify checks among angels to build a supportive network; choose early backers who create positive signals for later rounds; and avoid detractors by backchanneling with founders of failed ventures—ensuring investors add strategic value beyond capital. Jason Shuman’s conversation with Dan Shipper surfaces key principles for AI-native organizations: the shift from a knowledge economy to an “allocation economy” where orchestration of human and machine intelligence is paramount; the resurgence of generalists with strong taste and direction; and “compound engineering,” capturing prompt lessons to improve AI agents over time. AI Industry Developments & News Guillermo Rauch spotlights OpenAI’s GPT-5.2 Pro working with Harmonic to near-autonomously generate a proof for an Erdős mathematical problem—demonstrating how advanced language models are tackling complex reasoning tasks once reserved for human experts.
Related
A coding environment for Claude mentioned for its keyboard shortcut that opens a full-featured editor for prompt writing. It is highlighted as making long prompts far easier to manage.
Founder and operator referenced in a conversation about AI-native organizations. For PMs, he is associated with product thinking around orchestration, generalists, and AI-native companies.
A creator who demonstrates the Compound Engineering plugin and Claude Code workflow patterns.
Stay updated on Compound Engineering
Get curated AI PM insights delivered daily — covering this and 1,000+ other sources.
Subscribe Free