Keep context small. Drop stale, keep task state.
Rules
- Summarize old tool output into 1-2 lines; keep only the result that matters.
- Drop: resolved errors, obsolete file reads, duplicate logs.
- Keep: current plan, open TODO, key IDs/paths, last decision.
- Use session_search / notes for history instead of re-pasting.
- Prefer code/script over long prose explanations in context.
- Pre-thinking compression (biggest savings): BEFORE the model reasons, compress retrieved docs / RAG context / long file dumps into key facts + a source pointer. Drop full-doc dumps.
- Effort pinning: trivial task -> signal low effort, no over-think. Reuse a prior conclusion instead of re-reasoning the same thing.
Scope boundary (don't collide with siblings)
- Context hygiene + effort pinning + pre-thinking compression = THIS skill (runs automatically during the session).
paleo-converse= explicit conversation-turn compression + dedup (on-demand, when the user points at history or context is heavy).paleo-summary= one bulky artifact (log / diff / dump / stdout) -> tight intisari.- Hand off to those when the target is a specific turn-block or a single artifact; don't re-implement them here.
When
- Long session (>20 turns), big file dumps, repeated similar output.
- Before calling an expensive model — shrink first.
Gotchas
- Don't trim the user's latest instruction.
- Keep ground truth (file paths, IDs) — summarize derivations only.
- If unsure what's stale, ask once. Don't guess-drop critical state.