← sources

Critique my setup please

source: reddit (reddit:ClaudeAI)published 2026-10-25 00:28ingested 2026-10-26 02:36

retained text

I'm not a software engineer. I've built a setup where Claude does nearly all the coding and running of a few small websites, and I'd like people who've done this longer to tell me what's dumb about it. This is only about the system, not what the sites are.

\*\*The machine\*\*

One Mac mini at home. All the code, one SQLite database, the scheduled jobs and the site builds live there. When a site changes, the mini builds it and pushes it to a static host. The only things outside the mini are the hosting and a small serverless worker for sign-ins.

\*\*How I talk to it\*\*

Claude Projects. Each project has a main chat with a "coordinator" Claude. When I ask for something, it spins up a separate thread for that job. The thread connects to the Mac mini through Claude Code's Remote Control, does the work there, and reports back.

So I never open a terminal. I type in a thread, and Claude Code runs on my Mac.

Two Claude Code sessions also stay running on the mini. One does day-to-day work. The other is the only one allowed to change anything the sites show as a calculated number. Other threads can propose those changes but not make them.

\*\*Scheduling\*\*

Plain macOS launchd. Around 20 jobs, hourly to weekly. Jobs that can't overlap take a file lock.

\*\*Git is my knowledge base. I have a repo for every project and decisions and rules are stored there. 

Every decision I make gets a number and its own file in the repo. Decisions are never edited. To change one, you write a new one that says which it replaces. A script turns all of them into one generated "current rules" file, CI checks it, and every session reads it first. I'm at about 200.

There's a second repo of shared standards, about 17 principles that go across all projects, that every project copies read-only. A project can override one only by writing a decision that names it.

The rest of the working state is four plain files in the repo: a one-page handoff note each session overwrites, a bug log, a backlog, and a questions file where every question must come with a proposed answer.

\*\*Bugs and audits\*\*

Every bug goes in the bug log the same day it's found, tagged with one of a handful of categories. Each fix gets a review in a separate session: what caused it, what guard stops it coming back, and whether the same bug exists anywhere else. When a category keeps repeating, it becomes a test or a warning whose message says how to fix it. Guards start as warnings, not hard stops.

On top of that, a health check runs every 6 hours, a self-test runs the full test suite every night, and a monthly audit reviews memory for stale or contradicting facts.

\*\*Memory\*\*

Claude's project memory is a folder of small files with an index. It holds preferences and pointers, never rules. If memory and the repo disagree, the repo wins.

I’m gonna have a memory “dreaming” session once a week where fable organizes and checks the weeks memory work.  

\*\*What the agents can do on their own\*\*

Three tiers:

\- Do it and log it: bug fixes with tests passing, schedule tweaks, docs.  
\- Do it and tell me: bigger changes go in one nightly email. Nothing on quiet days. I can reverse anything.  
\- Ask me first: anything public, deletes, spending, credentials, or anything that can't be undone.

Unattended runs have caps on rounds, time and cost. Anything fetched from outside (web pages, emails) is treated as data, never instructions.

I also have a board of Grok agents set up as a sort of advisory board. Claude talks to one of them through files in a shared cloud folder.  They are either brain to bounce ideas off of like revenue and marketing ideas.  I like Claude and grok vetting with 2 brains. 

Critique everything. Suggest better ways. I'm especially curious how the decision and bug setup looks to people who've run something like this. The goal is to have a self fixing and self grading system that is also self learning. 

document fetched 2026-10-26 00:23 · canonical https://reddit.com/r/ClaudeAI/comments/1xeo991/critique_my_setup_please