How to Work and Compound with AI

How to Work and Compound with AI is a long-form guide by Ziyou (Eugene) Yan, published on eugeneyan.com in May 2026, on setting up a personal workflow with coding agents so that each finished task improves the next one. The post is organized around five principles, each with its own section: context as infrastructure, taste as configuration, verification for autonomy, scaling via delegation, and closing the loop. The examples are specific to Claude Code. Yan keeps code in ~/src and knowledge work in ~/vault so the model can find files with grep and glob, maintains a per-project INDEX.md that annotates each document with its URL, owner and when to read it, and treats a global ~/.claude/CLAUDE.md as a behavioral contract. Long CLAUDE.md sections move into guides that load lazily, and tasks done weekly or more become skills such as /polish, /write and /daily, bootstrapped by doing a task once and asking the model to codify it. The verification section describes a ladder that starts with post-edit hooks running ruff format and ruff check, then moves up through tests, evals and LLM review. The delegation section covers running three to six sessions in parallel in git worktrees and making them observable through stop hooks, tmux window titles and the Claude Code status line. The final section shows how to mine past transcripts for correction phrases such as "still wrong" and fold the fixes back into CLAUDE.md, skills or checks. A linked SETUP.txt lets readers have their own model apply the setup.

Topic

Skills

Level

Intermediate

Format

Guide

Cost

Free

Duration

~13 min read

Provider

Eugene Yan

Hands-on

Yes — code/exercises

Certificate

None

• ✓How to lay out code and notes so an agent can find context with grep and glob
• ✓How to write a per-project INDEX.md that tells the model which documents to read and when
• ✓How to scope CLAUDE.md files globally, per repository and per project directory
• ✓How to turn tasks you repeat weekly into small Claude Code skills with separate templates and scripts
• ✓How to build a verification ladder from post-edit lint hooks up to tests, evals and LLM review
• ✓How to run several agent sessions in parallel with git worktrees and keep them observable
• ✓How to mine past transcripts for correction phrases and turn them into configuration updates

• •Concrete file layouts, skill names and hook commands you can copy into your own setup
• •Ties how much work you can delegate to how cheaply you can verify the result
• •Includes a linked SETUP.txt that your own model can read to bootstrap the workflow
• •Covers knowledge work as well as code, with MCP connections to Slack, Drive and Mail

This page was written from 3 sources, 2 on domains other than eugeneyan.com.

1. 1.eugeneyan.com — working with aivendor
2. 2.news.ycombinator.com — item
3. 3.hn.algolia.com — search