Context Compression
3-layer pipeline for long sessions. From guanyang/antigravity-skills context-compression / agent-skills-for-context-engineering.
When to use
- Conversation exceeds ~50 turns or you notice context loss
- User says "compress", "compact", "too long"
- Before a complex multi-step task that needs headroom
Workflow
Use tokens-per-task, not tokens-per-request — a saving that causes re-fetching costs more.
Three layers:
- MicroCompact (30-40%, zero loss): Collapse confirmations, repeated tool outputs, filler.
- AutoCompact (60-70%, minimal loss): Restructure into
## What We Know, ## Decisions Made, ## Current Task; group file reads into summaries.
- Full Compact (85-90%, moderate loss but recoverable): Reduce to
## State, ## Intent, ## Constraints — no snippets, just paths.
Rules: Never compress away decisions, arch choices, active task state; always compress greetings, superseded attempts; flag if about to drop something critical.
Verification
References
- Source:
guanyang/antigravity-skills context-compression, irfad7/claude-power-skills context-compression
- See also:
productivity/token-efficiency for per-turn savings, productivity/prompt-enhancement for input
1---2name: context-compression3description: Use when agent sessions generate long histories, context window limits are hit, or you need to preserve decisions and artifact trails across compression.4---56# Context Compression783-layer pipeline for long sessions. From `guanyang/antigravity-skills` `context-compression` / `agent-skills-for-context-engineering`.910## When to use1112- Conversation exceeds ~50 turns or you notice context loss13- User says "compress", "compact", "too long"14- Before a complex multi-step task that needs headroom1516## Workflow1718Use **tokens-per-task**, not tokens-per-request — a saving that causes re-fetching costs more.1920**Three layers:**2122- **MicroCompact (30-40%, zero loss):** Collapse confirmations, repeated tool outputs, filler.23- **AutoCompact (60-70%, minimal loss):** Restructure into `## What We Know`, `## Decisions Made`, `## Current Task`; group file reads into summaries.24- **Full Compact (85-90%, moderate loss but recoverable):** Reduce to `## State`, `## Intent`, `## Constraints` — no snippets, just paths.2526**Rules:** Never compress away decisions, arch choices, active task state; always compress greetings, superseded attempts; flag if about to drop something critical.2728## Verification2930- [ ] Correct layer chosen for situation (getting long → Micro, losing context → Auto, hitting limits → Full)31- [ ] Output lists before/after estimate, preserved/dropped, and asks "Does this capture everything?"32- [ ] Artifact trail (files modified) preserved in structured sections3334## References3536- Source: `guanyang/antigravity-skills` context-compression, `irfad7/claude-power-skills` context-compression37- See also: `productivity/token-efficiency` for per-turn savings, `productivity/prompt-enhancement` for input
viasrijan