Not everyone needs this to use the system: the skills maintain the shape for you. It's here for when you want to understand the design, avoid the sharp edges, or run more than one project.
How it works introduced the two note types; two details matter once you're living in them.
The same fact rarely belongs in both stores: a decision is state; the task that acted on it is planning. Keeping them separate is what lets /status answer "what's left to do?" without wading through everything you've learned.
Every item is its own small markdown file, and a script regenerates INDEX.md from them; you never edit the index by hand. That regenerated index is what the LLM scans token-efficiently: it reads the index, opens only the item that matters, and reaches for a worklog only when it needs the full story.
/end-session can miss a state item buried in a file it didn't open. Review what it suggests; add the ones it didn't catch.
Worklog length drives how long /end-session takes to process. Keep entries terse; that also makes future /catchup sharper.
Meeting transcripts flip decision polarity and mangle names. Always read /process-meeting proposals before confirming.
One install can serve more than one project. meta.md supports a multi-project block, and every project-scoped skill takes an optional project name as its first argument (e.g. /status my-other-project) to pick which one it acts on. Re-run /setup to add a project block. For most people, though, a per-project clone is simpler: independent upgrades, no ambiguity about which project a bare command means.