AgentSkillsCN

Compress

压缩

SKILL.md

Super Compress

An extremely thorough context extraction followed by double compaction. Nothing is lost.

What This Skill Does

This is the "nuclear option" for context management. It extracts EVERYTHING valuable from the current conversation context, saves it to memory, then performs aggressive double compaction.

Phase 1: Deep Extraction (Before Any Compaction)

Extract EVERYTHING valuable from current context:

  1. •

    Decisions Made

    • •What was decided
    • •Why (full rationale with alternatives considered)
    • •Who/what influenced the decision
    • •Any tradeoffs accepted
  2. •

    Code Changes

    • •Files modified and why
    • •Key functions added/changed
    • •Patterns established
    • •Architecture decisions embedded in code
  3. •

    Errors & Debugging

    • •What broke
    • •Root cause analysis
    • •How it was fixed
    • •Prevention strategies learned
  4. •

    Discoveries

    • •Codebase insights
    • •Architecture learnings
    • •"Aha" moments
    • •Hidden dependencies found
  5. •

    Open Questions

    • •Unresolved issues
    • •Things to investigate later
    • •Blocked items and why
    • •Assumptions that need validation
  6. •

    Context for Next Session

    • •What we were working on
    • •Where we left off exactly
    • •Next steps planned
    • •Files that were open/relevant

Phase 2: Save to Memory

Use MCP memory tools to persist everything:

  • •Create new entities for major topics discovered
  • •Add detailed observations with evidence
  • •Create relationships between entities
  • •Tag observations with timestamps

Phase 3: Double Compact

  1. •First compact - reduce context by ~50%
  2. •Verify memory was saved correctly (read back key items)
  3. •Second compact - reduce to minimal working context

When to Use

  • •Context getting very long (>50k tokens)
  • •Before switching to a completely different task
  • •End of major feature work
  • •When you want a "fresh start" but keep all learnings
  • •Before ending a long session

Output

After compression:

code
=== SUPER COMPRESS COMPLETE ===

EXTRACTED:
- 3 decisions (saved to Decision entities)
- 5 code changes (saved to Session Log)
- 2 bugs fixed (saved to Bug Fix entities)
- 4 discoveries (saved to Architecture entity)
- 2 open questions (saved to Open Questions)

MEMORY WRITES:
✓ Created "Session 2025-12-30: Memory System Implementation"
✓ Added 8 observations to existing entities
✓ Created 2 new relationships

COMPACTION:
- Before: ~75k tokens
- After first compact: ~35k tokens
- After second compact: ~15k tokens

CONTEXT PRESERVED:
- Working on: Memory management system implementation
- Last completed: settings.json hooks
- Next: Create skill files
- Key files: .claude/settings.json, .claude/skills/

Ready to continue. All context saved to memory.

Process

  1. •Announce: "Starting super compress - extracting all context..."
  2. •Extract: Go through conversation systematically
  3. •Categorize: Sort into the 6 categories above
  4. •Save: Use mcp__memory__* tools to persist
  5. •Verify: Read back critical items
  6. •Compact: Request compaction (user may need to trigger)
  7. •Summarize: Show what was saved and new context size