// Compress agents/skills with quality assurance loop. Iterative 5-phase process ensures understanding preserved. TRIGGERS: semantic-compressor, compress skill, optimize agent, reduce tokens, benchmark
| name | semantic-compressor |
| description | Compress agents/skills with quality assurance loop. Iterative 5-phase process ensures understanding preserved. TRIGGERS: semantic-compressor, compress skill, optimize agent, reduce tokens, benchmark |
| user-invocable | true |
Compress agents/skills while preserving understanding. Uses 5-phase iterative quality assurance.
For EACH file to compress, execute ALL phases in order:
Goal: Confirm Claude understands the original file correctly.
# Extract understanding from original
claude -p "Read this and answer:
1. What is the main purpose? (1 sentence)
2. What are the key features? (bullet list)
3. What commands/tools are available?
4. What triggers should activate this?
<file content here>
Output as structured YAML."
Save output as <name>.original-understanding.yaml
Goal: Apply compression rules to create compressed version.
⚠️ MUST PRESERVE Checklist (67% of files needed iteration due to missing these)
BEFORE compressing, extract and LIST these from original:
□ ALL triggers (every keyword must be preserved)
□ ALL commands (every command must be listed)
□ ALL tools (every tool must be listed)
□ ALL supported formats/types (file formats, output formats, protocols)
□ ALL feature categories (capabilities must not be omitted)
□ ALL database/language names (specific names cannot be replaced with "etc.")
Compression Rules:
FOR descriptions (frontmatter):
- Remove: "This skill/agent provides...", "comprehensive", "advanced"
- Remove: "Use this when...", "is designed to", "enables users to"
- Keep: ALL trigger keywords, file extensions, action verbs
- Format: [Purpose <15 words]. [Features as list]. TRIGGERS: [keywords]
- Target: 3:1 ratio
FOR body content:
- Paragraphs → single sentences
- Verbose lists → terse bullet points
- Multiple examples → best single example
- Keep: code blocks, tables (condensed), commands
- ⚠️ NEVER use "etc.", "and more", "such as" - list ALL items
- Target: 2:1 ratio
FOR agents:
- Description: 1-2 sentences max
- Tools: list ALL, no descriptions
- Process: numbered one-liners
- Output: template only
- Target: 2:1 ratio
Phase 2 Checklist (pre-compression verification)
Before writing compressed version, verify:
✓ Every trigger from original appears in compressed
✓ Every command from original appears in compressed
✓ Every tool from original appears in compressed
✓ Every feature category is mentioned (can be brief)
✓ All specific names (DBs, languages, formats) are listed
Write compressed version to <name>.compressed.md
Goal: Confirm Claude understands the compressed version correctly.
# Extract understanding from compressed
claude -p "Read this and answer:
1. What is the main purpose? (1 sentence)
2. What are the key features? (bullet list)
3. What commands/tools are available?
4. What triggers should activate this?
<compressed content here>
Output as structured YAML."
Save output as <name>.compressed-understanding.yaml
Goal: Compare original vs compressed understanding.
claude -p "Compare these two understanding extracts.
ORIGINAL:
<original-understanding.yaml>
COMPRESSED:
<compressed-understanding.yaml>
Questions:
1. Is the purpose identical? (yes/no)
2. Are all key features preserved? (list any missing)
3. Are all commands/tools preserved? (list any missing)
4. Are all triggers preserved? (list any missing)
5. Overall: PASS or NEEDS_IMPROVEMENT?
If NEEDS_IMPROVEMENT, specify what to add back."
Decision:
Goal: Generate final comparison report.
## <name>
| Metric | Original | Compressed | Ratio |
|--------|----------|------------|-------|
| Lines | X | Y | X:Y |
| Tokens (est.) | X | Y | X:Y |
### Understanding Comparison
- Purpose: ✓ Preserved
- Features: ✓ All preserved / × Missing: [list]
- Commands: ✓ All preserved / × Missing: [list]
- Triggers: ✓ All preserved / × Missing: [list]
### Quality: PASS ✓ / FAIL ×
| Command | Action |
|---|---|
/semantic-compressor | Auto-scan project .claude/skills/ and .claude/agents/, compress all |
/semantic-compressor <file> | Run 5-phase loop on single file |
/semantic-compressor benchmark | Run 5-phase loop on all samples |
/semantic-compressor verify <file> | Phase 1 + 3 only (no compression) |
When running /semantic-compressor without arguments:
# Find all skill and agent files in project
find .claude/skills -name "*.md" -o -name "SKILL.md"
find .claude/agents -name "*.md"
# Create timestamped backup before any changes
BACKUP_DIR=".claude/backups/$(date +%Y%m%d_%H%M%S)"
mkdir -p "$BACKUP_DIR"
cp -r .claude/skills "$BACKUP_DIR/"
cp -r .claude/agents "$BACKUP_DIR/"
echo "✓ Backup created: $BACKUP_DIR"
For each discovered file:
# Already compressed comment)# Compressed by semantic-compressor on <date> header## Compression Report
| File | Original | Compressed | Ratio | Iterations | Status |
|------|----------|------------|-------|------------|--------|
| skill-a.md | 500 | 120 | 4.2:1 | 1 | ✓ PASS |
| agent-b.md | 300 | 80 | 3.8:1 | 2 | ✓ PASS |
Total: X files compressed, Y tokens saved
Backup: .claude/backups/YYYYMMDD_HHMMSS/
Located in .claude/skills/semantic-compressor/samples/:
samples/
├── skills/
│ ├── verbose/ # Original verbose versions
│ └── compressed/ # Compressed versions
└── agents/
├── verbose/ # Original verbose versions
└── compressed/ # Compressed versions
| Type | Min Ratio | Max Info Loss |
|---|---|---|
| Skill description | 2.5:1 | 0% triggers |
| Skill body | 2:1 | 0% commands |
| Agent | 2:1 | 0% tools |
If ratio < threshold OR info loss > 0%, mark as NEEDS_IMPROVEMENT.