Command
Last session · Meta learnings · Charts
Last Session — 2026-03-27
- Paperclip local instance at 127.0.0.1:3100 — confirmed running
- Plan mode used — Best of Both Combined strategy — all 20 validation checks passed
Meta~0.8 hrs PO~11x compression
Meta Learnings — 2026-03-24
2026-03-24
API rate limits mid-session: subagent delegation protects main context but if a subagent hits the cap, work stops until the window resets — schedule long rebuilds at session start, not end.
Static JSON data layer (sessions.json) + sync script (wrap_update.py) removes all hand-editing of HTML — a wrap now becomes a CLI call, not a code change.
Tool exists ≠ tool is used. Build the tool AND update the protocol that invokes it or the efficiency gain never lands.
2026-03-23
HTTP 200 on write ≠ data is readable. Always read-back after storage layer changes.
Vercel Blob free tier: cumulative ops limit — deleting blobs doesn't unsuspend. Migrate early.
2026-03-22
Vercel stale content = stale alias, not CDN cache. Re-alias after every vercel --prod.
The gap between Claude Code and other AI tools is context. Memory + CLAUDE.md = instant grounding.
2026-03-21
Token cost is not just what you type — hooks inject full trigger files before you send a word.
Session log in a dashboard should be git log-driven — never rely on memory alone.
Charts
My Time — Session Duration by Day
Per-Session Bar + Tokens
Compression — PO vs 3-Person Team