Run several projects through an agent that reads a markdown hub first

Shinyenyi started a git repository on 16 September 2026 that holds no code, only markdown: a 3,100-word master context with standing rules, a sync log, a backlog, one file per project and the content of a personal website. An AI agent reads these before working, then uses tools to read Shinyenyi's other repositories, run builds and write files. Shinyenyi reviews and applies what it hands back, and the agent cannot push. Each code repo has a CLAUDE.md pointing at the hub. The website builds from the hub's site folder, and a habit tracker reads journal files from it every 15 minutes. The post does not name the agent product.

After 72 commits the sync log has 31 entries and the master context holds 13 rules, each written after something broke. A session dated itself a day early, long files came back cut off, a replacement page without front matter took two articles offline, and a private project file landed in the public website folder and stayed there five days. Status files for two projects lagged the real work for days.

Shinyenyi cannot see whether the agent read a file, only what it produced, and nothing checks the rules unless a build or test happens to. The hub is public, so private notes need a separate encrypted journal. Rules copied into each repo's CLAUDE.md can drift from the hub.