Context Rot Fix

Context rot is real. The longer a Claude or Codex session runs, the more it forgets, contradicts itself, re-argues settled decisions and quietly changes your numbers. Compacting squeezes the history, but the reasons behind your decisions leak away each time. Plenty of free skills explain context rot. This one fixes it. How it works Your project's memory moves out of the chat and into three small files that the agent keeps up to date: - CONTEXT.md: what does not change (goals, people, where things live, settled decisions, the official numbers) - STATE_NOW.md: where you are today (open loops with days open, next actions, what must not be done) - LEDGER.md: an append-only log of what was done, decided and checked Say "checkpoint" before a session degrades. Open a fresh chat and say "resume". The agent reads only those files, so the new context starts small and clean. It runs a freshness check to make sure your notes are not older than your work, reports where you are in 15 lines, and gets to work. What you get - The 5 rules that stop context rot - Resume and checkpoint routines the agent follows on command - Copy-ready templates for all three files - A freshness-check script that flags stale notes before they mislead the agent - Works for any multi-week project: code, research, writing, analysis, client work Fresh context, full memory. One SKILL.md for Claude Code, Claude apps, Codex and any agent that loads skills.