Sensei

"A true master teaches not by telling, but by refining."

Automates skill frontmatter improvement using the Ralph loop pattern - iteratively improving skills until they reach Medium-High compliance with passing tests.

Help

When user says "sensei help" or asks how to use sensei:

╔══════════════════════════════════════════════════════════════════╗
║  SENSEI - Skill Frontmatter Compliance Improver                  ║
╠══════════════════════════════════════════════════════════════════╣
║                                                                  ║
║  USAGE:                                                          ║
║    Run sensei on <skill-name>              # Single skill        ║
║    Run sensei on <skill-name> --fast       # Skip tests          ║
║    Run sensei on <skill1>, <skill2>        # Multiple skills     ║
║    Run sensei on all Low-adherence skills  # Batch by score      ║
║    Run sensei on all skills                # All skills          ║
║                                                                  ║
║  WHAT IT DOES:                                                   ║
║    1. READ    - Load skill's SKILL.md and count tokens           ║
║    2. SCORE   - Check compliance (Low/Medium/Medium-High/High)   ║
║    3. SCAFFOLD- Create tests from template if missing            ║
║    4. IMPROVE - Add WHEN: triggers (cross-model optimized)       ║
║    5. TEST    - Run tests, fix if needed                         ║
║    6. TOKENS  - Check token budget                               ║
║    7. SUMMARY - Show before/after comparison                     ║
║    8. PROMPT  - Ask: Commit, Create Issue, or Skip?              ║
║    9. REPEAT  - Until Medium-High score achieved                 ║
║                                                                  ║
║  TARGET SCORE: Medium-High                                       ║
║    ✓ Description > 150 chars, ≤ 60 words                         ║
║    ✓ Has "WHEN:" trigger phrases (preferred)                     ║
║    ✓ No "DO NOT USE FOR:" (risky in multi-skill envs)             ║
║    ✓ Has "INVOKES:" for tool relationships (optional)            ║
║    ✓ SKILL.md < 500 tokens (soft limit)                          ║
║                                                                  ║
║  MCP INTEGRATION (when INVOKES present):                         ║
║    ✓ Has "MCP Tools Used" table                                  ║
║    ✓ Has Prerequisites section                                   ║
║    ✓ Has CLI fallback pattern                                    ║
║    ✓ No skill-tool name collision                                ║
║                                                                  ║
╚══════════════════════════════════════════════════════════════════╝

Configuration

Sensei uses these defaults (override by specifying in your prompt):

Auto-detect skills directory by checking (in order):

1. skills/ in project root
2. .github/skills/
3. User-specified path

Invocation Modes

Single Skill

Run sensei on my-skill-name

Multiple Skills

Run sensei on skill-a, skill-b, skill-c

By Adherence Level

Run sensei on all Low-adherence skills
Run sensei on all Medium-adherence skills

All Skills

Run sensei on all skills

Fast Mode (Skip Tests)

Run sensei on my-skill --fast

The Ralph Loop

For each skill, execute this loop until score >= Medium-High:

Step 1: READ

Load the skill's current state:

{skills-dir}/{skill-name}/SKILL.md
{tests-dir}/{skill-name}/ (if exists)

Run token count:

npm run tokens -- count {skills-dir}/{skill-name}/SKILL.md

Step 2: SCORE

Assess compliance by checking the frontmatter for:

• Description length (>= 150 chars, ≤ 60 words)
• "WHEN:" trigger phrases (preferred) or "USE FOR:"
• Routing clarity ("INVOKES:", "FOR SINGLE OPERATIONS:")
• No "DO NOT USE FOR:" anti-triggers (risky in multi-skill environments)

See references/scoring.md for detailed criteria.

Step 3: CHECK

If score >= Medium-High AND tests pass → go to SUMMARY step.

Step 4: SCAFFOLD (if needed)

If {tests-dir}/{skill-name}/ doesn't exist, create test scaffolding using templates from references/test-templates/.

Step 5: IMPROVE FRONTMATTER

Enhance the SKILL.md description to include:

1. Lead with action verb - First sentence: unique action verb + domain
2. Trigger phrases - "WHEN:" (preferred) or "USE FOR:" with 3-5 distinctive quoted phrases
3. Keep description under 60 words and 1024 characters

⚠️ "DO NOT USE FOR:" carries context-dependent risk. In multi-skill environments (10+ skills with overlapping domains), anti-trigger clauses introduce the very keywords that cause wrong-skill activation on Claude Sonnet and fast-pattern-matching models (evidence). For small, isolated skill sets (1-5 skills), the risk is low. When in doubt, use positive routing with WHEN: and distinctive quoted phrases.

Template (cross-model optimized):

---
name: skill-name
description: "[ACTION VERB] [UNIQUE_DOMAIN]. [One clarifying sentence]. WHEN: \"[phrase1]\", \"[phrase2]\", \"[phrase3]\", \"[phrase4]\", \"[phrase5]\"."
---

Template (with routing clarity for High score):

---
name: skill-name
description: "**WORKFLOW SKILL** — [ACTION VERB] [UNIQUE_DOMAIN]. [Clarifying sentence]. WHEN: \"[phrase1]\", \"[phrase2]\", \"[phrase3]\". INVOKES: [tools/MCP servers used]. FOR SINGLE OPERATIONS: [when to bypass this skill]."
---

Step 6: IMPROVE TESTS

Update test prompts to match new frontmatter:

• shouldTriggerPrompts - 5+ prompts matching "WHEN:" or "USE FOR:" phrases
• shouldNotTriggerPrompts - 5+ prompts for unrelated topics and different-skill scenarios

Step 7: VERIFY

Run tests (skip if --fast flag):

# Framework-specific command based on project
npm test -- --testPathPattern={skill-name}  # Jest
pytest tests/{skill-name}/                   # pytest
waza run tests/{skill-name}/trigger_tests.yaml  # Waza

Step 8: TOKENS

Check token budget:

npm run tokens -- check {skills-dir}/{skill-name}/SKILL.md

Budget guidelines:

• SKILL.md: < 500 tokens (soft), < 5000 (hard)
• references/*.md: < 1000 tokens each

Step 8b: MCP INTEGRATION (if INVOKES present)

When description contains INVOKES:, check:

1. MCP Tools Used table - Does skill body have the table?
2. Prerequisites section - Are requirements documented?
3. CLI fallback - Is there a fallback when MCP unavailable?
4. Name collision - Does skill name match an MCP tool?

If checks fail, add missing sections using patterns from mcp-integration.md.

Step 9: SUMMARY

Display before/after comparison:

╔══════════════════════════════════════════════════════════════════╗
║  SENSEI SUMMARY: {skill-name}                                    ║
╠══════════════════════════════════════════════════════════════════╣
║  BEFORE                          AFTER                           ║
║  ──────                          ─────                           ║
║  Score: Low                      Score: Medium-High              ║
║  Tokens: 142                     Tokens: 385                     ║
║  Triggers: 0                     Triggers: 5                     ║
║  Anti-triggers: 0                Anti-triggers: 3                ║
╚══════════════════════════════════════════════════════════════════╝

Step 10: PROMPT USER

Ask how to proceed:

• [C] Commit - Save with message sensei: improve {skill-name} frontmatter
• [I] Create Issue - Open issue with summary and suggestions
• [S] Skip - Discard changes, move to next skill

Step 11: REPEAT or EXIT

• If score < Medium-High AND iterations < 5 → go to Step 2
• If iterations >= 5 → timeout, show summary, move to next skill

Scoring Quick Reference

⚠️ "DO NOT USE FOR:" is risky in multi-skill environments (10+ overlapping skills) — causes keyword contamination on fast-pattern-matching models. Safe for small, isolated skill sets. Use positive routing with WHEN: for cross-model safety.

MCP Integration Score (when INVOKES present)

See references/scoring.md for full criteria. See references/mcp-integration.md for MCP patterns.

Frontmatter Patterns

Skill Classification Prefix

Add a prefix to clarify the skill type:

• **WORKFLOW SKILL** - Multi-step orchestration
• **UTILITY SKILL** - Single-purpose helper
• **ANALYSIS SKILL** - Read-only analysis/reporting

Routing Clarity (for High score)

When skills interact with MCP tools or other skills, add:

• INVOKES: - What tools/skills this skill calls
• FOR SINGLE OPERATIONS: - When to bypass this skill

Quick Example

Before (Low):

description: 'Process PDF files'

After (High with routing, cross-model optimized):

description: "**WORKFLOW SKILL** — Extract, rotate, merge, and split PDF files. WHEN: \"extract PDF text\", \"rotate PDF pages\", \"merge PDFs\", \"split PDF\". INVOKES: pdf-tools MCP for extraction, file-system for I/O. FOR SINGLE OPERATIONS: Use pdf-tools MCP directly for simple extractions."

See references/examples.md for more before/after transformations.

Commit Messages

sensei: improve {skill-name} frontmatter

Reference Documentation

• scoring.md - Detailed scoring criteria and algorithm
• mcp-integration.md - MCP tool integration patterns
• loop.md - Ralph loop workflow details
• examples.md - Before/after transformation examples
• configuration.md - Project setup patterns
• test-templates/ - Test scaffolding templates
• test-templates/waza.md - Waza trigger test format

Built-in Scripts

Run npm run tokens help for full usage.

Token Commands

npm run tokens count              # Count all markdown files
npm run tokens check              # Check against token limits
npm run tokens suggest            # Get optimization suggestions
npm run tokens compare            # Compare with git history

Configuration

Create .token-limits.json to customize limits:

{
  "defaults": { "SKILL.md": 500, "references/**/*.md": 1000 },
  "overrides": { "README.md": 3000 }
}