claude-code-workspace-public

This is a sanitized public mirror of the real Claude Code workspace I (Rostyslav Khomyak) use day to day to run automation work across several clients. The private original holds live client data, so this copy has client names, directories, and domains genericized (ClientA, ClientB, …) and all secrets and incident specifics stripped. The structure, rules, skills, hooks, and self-learning loop are real and unedited except for those redactions.

What this shows

I don't paste prompts into a chat window and copy results out. Claude Code is the environment the work happens in: a terminal workspace with a layered CLAUDE.md, custom skills, hooks, a file-based memory, and a self-learning loop that turns repeated mistakes into hard rules.

How the CLAUDE.md got to its current state

It started as a routing file to keep clients isolated and grew into about a dozen sections. Most of its rules are scar tissue, not up-front design — they exist because I watched the model make the same mistake twice. The path is deliberate: a correction gets logged to _workspace/patterns/mistakes.jsonl; when a mistake repeats, it graduates into CLAUDE.md (or a client's own CLAUDE.md) as a hard rule, so the model is constrained instead of merely told. See _workspace/patterns/README.md for the loop, and _workspace/rules/self-learning.md for the promotion rules.

A few representative rules that earned their place this way: