What is Systemix?
Systemix runs little experiments on your website, writes down what it learned and why, and suggests the next one. You approve every decision.
That's the whole product. It runs in Claude Code, and everything it knows lives as plain files in your repo — no platform, no dashboard you don't own.
The four steps
- Propose — the daily runner suggests the next experiment as a card in
.systemix/queue.json, built from what the loop has already learned. It never creates anything itself — you say yes or no. - Build — the experiment becomes a file:
experiments/<id>.mdx. The bet, who it's for, and the number that will prove it. - Measure — the runner pulls real numbers (PostHog) into the file, every day, on its own.
- Learn — when the numbers are strong enough, you get a card recommending a decision. You close it; the lesson is written to
experiments/LEARNINGS.md— the notebook the next proposal starts from.
What you approve
The system moves on its own, but three things always wait for your OK:
- Starting a new experiment — the runner proposes;
/init-experimentis how you accept. - Closing an experiment — the runner recommends promote / iterate / kill; you decide.
- Changing its own skills or rules — always proposed, never self-applied.
Three ways to drive it
- Claude Code skills —
/systemix-init,/measure,/close-experimentand more - CLI —
systemix experiment new | list | close | learningsandsystemix loop - MCP —
experiment_new,experiment_list,experiment_closefor any agent that speaks MCP
All three doors read and write the same files.
Optional adapters — a design-system spine (design/, guardrails, tokens) and Figma sync ship in the box, but the loop doesn't need them. See the design-system adapter.
Start here
→ Getting started — set up the loop in a new or existing repo
→ The loop — each step in detail, and what the daily runner does
→ The experiment — the anatomy of one experiment file
→ Why files & loops — why the memory lives on disk
→ Skills reference — every skill, generated from the source