Semantic Compressor

Compress agents/skills while preserving understanding. Uses 5-phase iterative quality assurance.

Workflow: 5-Phase Quality Loop

For EACH file to compress, execute ALL phases in order:

Phase 1: Original Verification

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

Phase 2: Compress

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

Phase 3: Compressed Verification

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

Phase 4: Compare & Decide

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:

• If PASS → Phase 5 (Report)
• If NEEDS_IMPROVEMENT → Return to Phase 2 with feedback

Phase 5: Report

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 ×

Quick Commands

/semantic-compressor

.claude/skills/

.claude/agents/

/semantic-compressor <file>

/semantic-compressor benchmark

/semantic-compressor verify <file>

Project Directory Auto-Scan Mode

When running

/semantic-compressor

Step 1: Discovery

# Find all skill and agent files in project
find .claude/skills -name "*.md" -o -name "SKILL.md"
find .claude/agents -name "*.md"

Step 2: Backup (Required)

# 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"

Step 3: Process Each File

For each discovered file:

1. Skip files already compressed (check for

   # Already compressed

   comment)
2. Run 5-phase workflow
3. Replace original with compressed version (original in backup)
4. Add

   # Compressed by semantic-compressor on <date>

   header

Step 4: Generate Report

## 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/

Sample Files

Located in

.claude/skills/semantic-compressor/samples/

samples/
├── skills/
│   ├── verbose/        # Original verbose versions
│   └── compressed/     # Compressed versions
└── agents/
    ├── verbose/        # Original verbose versions
    └── compressed/     # Compressed versions

DO NOT Compress

• Error messages (need exact text)
• Security instructions
• Code syntax
• Numerical thresholds
• File paths

Quality Thresholds

If ratio < threshold OR info loss > 0%, mark as NEEDS_IMPROVEMENT.