When your agent sessions balloon past context limits or you're working with 5M+ token codebases, this is the compression playbook you need. It lays out three production strategies, with the standout being "anchored iterative summarization" that keeps structured summaries with dedicated sections for intent, file changes, and decisions. The key insight: optimize for tokens per task, not tokens per request, because aggressive compression that loses critical info just makes you refetch everything later. Especially useful when debugging why your agent forgot which files it modified. Originally from muratcankoylan's agent skills collection, now maintained by flora131.
npx skills add https://github.com/flora131/atomic --skill context-compression