Sensei

"A true master teaches not by telling, but by refining." - The Skill Sensei

Automates skill frontmatter improvement using the Ralph loop pattern - iteratively improving skills until they reach Medium-High compliance with passing tests, then checking token usage and prompting for action.

Help

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

╔══════════════════════════════════════════════════════════════════╗
║  SENSEI - Skill Frontmatter Compliance Improver                  ║
╠══════════════════════════════════════════════════════════════════╣
║                                                                  ║
║  USAGE:                                                          ║
║    Run sensei on <skill-name>              # Single skill        ║
║    Run sensei on <skill-name> --skip-integration  # Fast mode    ║
║    Run sensei on <skill1>, <skill2>, ...   # Multiple skills     ║
║    Run sensei on all Low-adherence skills  # Batch by score      ║
║    Run sensei on all skills                # All skills       ║
║                                                                  ║
║  EXAMPLES:                                                       ║
║    Run sensei on appinsights-instrumentation                     ║
║    Run sensei on azure-security --skip-integration               ║
║    Run sensei on azure-security, azure-observability             ║
║    Run sensei on all Low-adherence skills                        ║
║                                                                  ║
║  WHAT IT DOES:                                                   ║
║    1. READ      - Load skill's SKILL.md, tests, and token count  ║
║    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. REFERENCES- Validate markdown links                        ║
║    7. TOKENS    - Check token budget, gather suggestions         ║
║    8. SUMMARY   - Show before/after with suggestions             ║
║    9. PROMPT    - Ask: Commit, Create Issue, or Skip?            ║
║   10. REPEAT    - Until Medium-High score + tests pass           ║
║                                                                  ║
║  TARGET SCORE: Medium-High                                       ║
║    ✓ Description > 150 chars, ≤ 60 words                         ║
║    ✓ Has "WHEN:" trigger phrases (preferred)                     ║
║    ✓ No "DO NOT USE FOR:" (unless disambiguation-critical)         ║
║    ✓ SKILL.md < 500 tokens (soft limit)                          ║
║                                                                  ║
║  MORE INFO:                                                      ║
║    See .github/skills/sensei/README.md for full documentation    ║
║                                                                  ║
╚══════════════════════════════════════════════════════════════════╝

When to Use

• Improving a skill's frontmatter compliance score
• Adding trigger phrases and anti-triggers to skill descriptions
• Batch-improving multiple skills at once
• Auditing and fixing Low-adherence skills

Invocation Modes

Single Skill

Run sensei on azure-deploy

Multiple Skills

Run sensei on azure-security, azure-observability

By Adherence Level

Run sensei on all Low-adherence skills

All Skills

Run sensei on all skills

GEPA Mode (Deep Optimization)

Run sensei on my-skill --gepa
Run sensei on my-skill --gepa --skip-integration
Run sensei on all skills --gepa

When --gepa is used, Step 5 (IMPROVE) is replaced with GEPA evolutionary optimization. Instead of template-based improvements, GEPA parses trigger prompt arrays from the existing test harness and combines them with content quality heuristics to build a fitness function. An LLM proposes and evaluates many candidate improvements automatically. Note: GEPA does not execute Jest tests directly — it uses the test data (prompts) as evaluation inputs.

GEPA score-only mode (no LLM calls, just evaluate current quality):

Run sensei score my-skill
Run sensei score all skills

The Ralph Loop

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

1. READ - Load plugin/skills/{skill-name}/SKILL.md, tests, and token count
2. SCORE - Run spec-based compliance check (see SCORING.md):
   • Validate name per agentskills.io spec (no --, no start/end -, lowercase alphanumeric)
   • Check description length and word count (≤60 words)
   • Check triggers (WHEN: preferred, USE FOR: accepted)
   • Warn on "DO NOT USE FOR:" (risky in multi-skill environments — exception: REQUIRED for skills that share trigger overlap with broader skills like azure-prepare)
   • Preserve optional spec fields (license, metadata, allowed-tools) if present
3. CHECK - If score >= Medium-High AND tests pass → go to TOKENS step
4. SCAFFOLD - If tests/{skill-name}/ doesn't exist, create from tests/_template/
5. IMPROVE FRONTMATTER - Add WHEN: triggers (stay under 60 words and 1024 chars) 5b. IMPROVE WITH GEPA (when --gepa flag is set) — Replaces step 5 (IMPROVE FRONTMATTER) with automated optimization; step 6 (IMPROVE TESTS) still runs normally:
   • Auto-discovers tests/{skill-name}/triggers.test.ts and extracts prompt arrays
   • Builds a GEPA evaluator scoring content quality + trigger accuracy based on those trigger prompt arrays (not Jest test pass/fail results)
   • Runs python .github/skills/sensei/scripts/gepa/auto_evaluator.py optimize --skill {skill-name} --skills-dir plugin/skills --tests-dir tests
   • Shows diff of optimized SKILL.md for user approval
   • GEPA uses existing test trigger definitions as configuration — it does not execute, replace, or modify Jest tests
6. IMPROVE TESTS - Update shouldTriggerPrompts and shouldNotTriggerPrompts to match the finalized frontmatter (including any GEPA changes)
7. VERIFY - Run cd tests && npm test -- --testPathPatterns={skill-name}
8. VALIDATE REFERENCES - Run cd scripts && npm run references {skill-name} to check markdown links
9. TOKENS - Check token budget and line count (< 500 lines per spec), gather optimization suggestions
10. SUMMARY - Display before/after comparison with unimplemented suggestions
11. PROMPT - Ask user: Commit, Create Issue, or Skip?
12. REPEAT - Go to step 2 (max 5 iterations per skill)

Scoring Criteria (Quick Reference)

Sensei validates skills against the agentskills.io specification. See SCORING.md for full details.

Target: Medium-High (distinctive triggers, concise description)

⚠️ "DO NOT USE FOR:" is risky in multi-skill environments (15+ 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.

Exception — disambiguation-critical skills: When a skill's USE FOR triggers directly overlap with a broader skill (e.g., azure-prepare owns "deploy to Azure"), DO NOT USE FOR: is REQUIRED to prevent the broader skill from capturing prompts that belong to the specialized skill. Removing it causes routing regressions. Integration tests validate this routing -- run them before removing any DO NOT USE FOR: clause.

Strongly recommended (reported as suggestions if missing):

• license — identifies the license applied to the skill
• metadata.version — tracks the skill version for consumers

Frontmatter Template

Per the agentskills.io spec, required and optional fields:

---
name: skill-name
description: "[ACTION VERB] [UNIQUE_DOMAIN]. [One clarifying sentence]. WHEN: \"trigger 1\", \"trigger 2\", \"trigger 3\"."
license: MIT
metadata:
  version: "1.0"
# Other optional spec fields — preserve if already present:
# metadata.author: example-org
# allowed-tools: Bash(git:*) Read
---

IMPORTANT: Use inline double-quoted strings for descriptions. Do NOT use >- folded scalars (incompatible with skills.sh). Do NOT use | literal blocks (preserves newlines). Keep total description under 1024 characters and ≤60 words.

⚠️ "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.

Exception: DO NOT USE FOR: is REQUIRED when a specialized skill's triggers overlap with a broader skill (e.g., azure-hosted-copilot-sdk vs. azure-prepare on "deploy to Azure"). Without the negative discriminator, the broader skill captures prompts that should route to the specialized one. Always run integration tests before removing a DO NOT USE FOR: clause.

Test Scaffolding

When tests don't exist, scaffold from tests/_template/:

cp -r tests/_template tests/{skill-name}

Then update:

1. SKILL_NAME constant in all test files
2. shouldTriggerPrompts - 5+ prompts matching new frontmatter triggers
3. shouldNotTriggerPrompts - 5+ prompts matching anti-triggers

Commit Messages:

sensei: improve {skill-name} frontmatter

Constraints

• Only modify plugin/skills/ - these are the Azure skills used by Copilot
• .github/skills/ contains meta-skills like sensei for developer tooling
• Max 5 iterations per skill before moving on
• Description must stay under 1024 characters
• SKILL.md should stay under 500 tokens (soft limit)
• Tests must pass before prompting for action
• User chooses: Commit, Create Issue, or Skip after each skill

Flags

⚠️ Skipping integration tests speeds up the loop but may miss runtime issues. Consider running full tests before final commit.

Reference Documentation

• SCORING.md - Detailed scoring criteria
• LOOP.md - Ralph loop workflow details
• EXAMPLES.md - Before/after examples
• TOKEN-INTEGRATION.md - Token budget integration

Related Skills

• markdown-token-optimizer - Token analysis and optimization
• skill-authoring - Skill writing guidelines