Arty's Claude Newsletter
Escaped the labReally an experiment in Claude routines and how you lay out a repo for an AI to work in. It just happens to email me an opinionated AI newsletter every morning.
The newsletter is the excuse. The actual question was: how do Claude routines behave when you let one run unattended every day, and what does a repo have to look like for an AI to do good work inside it?
The answer the lab landed on: file and directory separation *is* the setup. Don’t hand a model one giant prompt — give it a directory where every kind of knowledge lives in its own file, and the boundaries do the thinking for you. Here that means an interests file for taste, a rubric for scoring, a rolling feedback file that learns what to boost and drop, a template for the output shape, a sources list, and an archive of past issues for continuity. Each one is separately editable, separately debuggable, and separately swappable. Change the taste without touching the format; change the format without touching the sources.
The other half of the discipline is the split between code and judgment: Python does everything deterministic — fetch, dedupe, hash, persist state — and Claude does everything that needs an opinion. It scores each candidate against the rubric, discards ruthlessly, clusters survivors into themes, tracks which ones continue running “threads” from earlier issues, drafts, then critiques its own draft (real takes or disguised summaries? filler? weak subject line?) and rewrites before anything ships. It emails the issue, commits it to git as the archive, and updates state so tomorrow knows what today said.
It runs as a scheduled routine with no server and no API bill. But the transferable part isn’t the newsletter — it’s the directory layout.