Post

I'm building a new product end to end through Slack. No IDE open. No terminal. My dev environment is a chat thread with an AI agent. Here is the whole stack: • Slack • Hermes (on Opalstack) • OpenCode Go Sharing my experience here

I host Hermes on OpalStack I wrote about my setup here: www.jjude.com/tech-notes/... Opalstack has been my VPS for many years. All my apps are hosted there. They now support podman too. It became the default choice for me.

Why Slack? Because with slack, I don't have to expose any port from opalstack. Hermes uses Socket Mode mode to connect to slack So no public exposure of Hermes It doesn't eliminate security threats; but reduces it to a large extent.

Everything lives in a wiki I have been using a framework which I named WINS: • wealth • insights • network • self My LLM wiki also follows that structure. I wrote about it here: www.jjude.com/tech-notes/...

All the product code lives under produce/hermes/products When I start coding a product, hermes creates a new folder under this main folder. Everything related to the products goes here.

Research PRD Mockups Feature Breakups Code Tasks for generating each features Retro of each cycle

When I have a "thinking" job, I sit on my computer and chat with the agent (via slack). Thinking Job = what is the product, JTBD, what it shouldn't be etc. This goes back and forth Often this will have >150 messages or so

This results in a PRD and other documents. At the end of a thread, I will ask the agent to document as well as to create a hand-off message. With a hand-off message, I move to a new thread. (After 100 - 120 messages the cost goes up and possibility of drift also goes up)

When "doing" comes, I am not so attached to my laptop. I pick up my mobile and instruct the agent as I move around. I ask it to create a task.md for each feature; I ask it to be sure of each task item. Then I ask it to generate code, while I am busy elsewhere.

After it generates code, it runs tests. If the tests pass it will tell me to test the code. Then I will test it and give feedback (again via slack). 90% of them this goes ok Some times the feature doesn't work as I expect.

What surprised me in all these: My time goes to deciding, reviewing plans, and saying no. The agent does the mechanical work and catches what I miss.