agent-workspace

One entry point for everything you're working on. A root file points your agent to every project; each project carries its own index. The agent reaches all your work but only loads the slice it needs — so it stays sharp as you pile more in, instead of getting dumber as its context fills up (context rot).

Three rules are the whole idea:

• Index, don't dump. The root CLAUDE.md lists projects and points to them — never their contents.
• Every project carries its own index. Each topics/ / has a CLAUDE.md the agent reads before the rest.
• Sweep on a cadence. The /update skill re-scans each working session and writes what actually happened — decisions, findings, status changes — into the right files, fixing anything now stale or contradictory. This is the part the toy "second brain" setups skip — and why they don't survive scale.

Two skills do the maintenance:

• **/update** — end-of-session capture sweep. Everything durable from the conversation lands in the right topic, the indexes stay true, and the latest discussion wins over stale notes.
• **/task** — for multi-session work with real stages. One living spec file per task in tasks/ (stage, plan, decision log — resumable cold), nudged through Backlog → Spec → In Progress → Code Review → Testing → Done. The spec file is the tracker; wire in Linear/Jira later if you want, the depth stays in the file.