samarv/compounding-engineering-with-ai

Compounding Engineering with AI

Compounding engineering is the practice of ensuring that every unit of work performed makes the next unit of work easier. Instead of just completing a task, you build a "simulation" of your decision-making process into an AI agent, creating permanent operational leverage for the team.

The Compounding Workflow

1. Identify the "Broken Record" Tasks

Look for any task where a senior team member or manager is repeating themselves. Use the "CEO Benchmark": if you have to say the same thing in a meeting twice, it should be codified into a prompt.

• Common triggers: Headline feedback, copy editing for style, PRD formatting, or code style reviews.

2. Codify Your "Taste"

AI models like Claude 3.5 Sonnet or Opus 4 can now act as "judges" of quality if given the right context.

• Record and Transcribe: Record your verbal feedback to a teammate using tools like Granola or Otter.
• Extract Principles: Feed the transcript into an LLM and ask: "Based on this feedback, what are my core principles for [Subject]?"
• Create the Style Guide: Turn those principles into a specific "System Prompt" that the AI can use to evaluate future work.

3. Build the "Simulation of Self" Agent

Create a standalone tool or prompt that others must use before they come to you for review.

• Input: Raw drafts or "rambling thoughts."
• Processing: The AI applies your codified taste/style guide.
• Output: A refined version that meets 80% of your requirements.
• Refinement: If the output is wrong, don't just fix it manually. Update the prompt/style guide so the mistake never happens again.

4. Implement Technical Speed-ups

For engineering or product teams, move beyond one-off chats into shared automation.

• Prompt Repositories: Store all high-value prompts in a shared GitHub repository or internal tool (e.g., Every uses a library of "Slash Commands" in Claude Code).
• PRD Generators: Build a prompt that takes raw audio or messy notes and outputs a perfectly formatted PRD based on your team's specific template.
• Automated Reviewers: Set up agents (like "Charlie" or "Friday") to live in your GitHub or Slack and automatically flag deviations from your team's specific standards.

Examples

Example 1: Editorial Style Automation

• Context: An Editor-in-Chief spends 2 hours daily fixing "punchiness" and tone in newsletters.
• Application: The Editor records 30 minutes of live feedback sessions. An AI Ops lead transforms this into an "Every Style Guide" prompt.
• Output: A "Copy-Editor Agent" that writers run their drafts through. The Editor now only reviews the final 10% of nuance, saving 90 minutes a day.

Example 2: The "Rambling to PRD" Loop

• Context: A Product Manager has great ideas but struggles to spend hours writing formal PRDs.
• Application: The PM records a 5-minute "brain dump" into a voice recorder. They feed this into a "Compounding PRD" prompt that knows their specific technical stack and formatting preferences.
• Output: A high-fidelity PRD ready for an AI Coder (like Claude Code) to begin implementation immediately.

Common Pitfalls

• The Behavioral Gap: Building the tool is easy; getting the team to use it is hard. You must make the AI-agent step a mandatory part of the workflow (e.g., "Don't send me a draft until the Style Agent has seen it").
• Static Prompts: Treating a prompt like a "set it and forget it" tool. Prompts need "context engineering"—regularly updating them with new examples of "Good" vs "Bad" outputs.
• Manual Fixing: Correcting an AI's mistake yourself instead of fixing the prompt. This kills the compounding effect. Always "fix the factory, not the product."
• Over-specialization: Hiring for narrow roles. In an allocation economy, you need generalists who can manage multiple agents across different domains (coding, writing, research).