Claude SOPs vs. io.github.datasynx/agentic-ai-shadowing

Keep your own standard operating procedures — the recurring things you do a particular way — as a private, versioned library of markdown files the agent reads on demand and follows. Distinct from sop-writer, which authors printed SOP documents for other people; these are procedures an agent executes. One delimited block in your CLAUDE.md points at the library, a generated INDEX.md carries id, title and a 'use when' trigger for each procedure, and exactly one file is read once a situation matches — so a library of forty procedures costs one table to consult rather than forty resident skill descriptions. Seven skills cover setup, run, write, edit, list, retire and sync. Procedures live in ~/.claude-user-data/sops/ as flat readable files inside a private git repo, mirrored across machines; the plugin repo never holds one. Each SOP declares an autonomy level (auto, confirm, confirm-each, manual) that records a default rather than granting permission, and a last_verified date meaning the procedure was seen to work

Observes tasks and generates anonymized SOPs — task tracking and SOP review via MCP. Fully local.