Six weeks of building, two environments, and 100 documented lessons. A developer who built a personal AI agent - not a chatbot wrapper, but a persistent assistant that reads emails, tracks deals, analyzes business data, and flags things you'd otherwise miss - shared the full breakdown in a detailed writeup.
The agent runs on Claude Projects, which lets you attach shared memory files - persistent documents Claude can read across every conversation. That's different from the standard context window, the amount of text a model can process in a single session. With Projects, the agent knows what happened yesterday because you've written it down in files Claude always has access to. The developer ran it across two environments - a local setup and a cloud deployment - and the lessons are split between what worked in each.
The distinction that keeps coming up in the writeup: most people using Claude stop at question-and-answer. This agent has state. It remembers which deals were open last week, which emails haven't been answered, which numbers looked off last Tuesday. That persistence is what separates a useful assistant from a sophisticated search box.
The 100-tip format covers the practical breakdown: how to structure memory files, when to give Claude explicit rules versus letting it reason, and how to handle the failure cases that cost the most time. Most AI agent tutorials cover the happy path. This one covers the six weeks of debugging that follow it.
For freelancers, small business owners, or developers who've considered building something similar and stopped because it seemed too complex - this kind of lived-experience documentation is harder to find than the tutorials.