Skill Index — Progressive Disclosure Engine

Save 90%+ tokens. Agents scan the index first, load full skills only when executing. Inspired by Loki Mode's 3-Layer Progressive Disclosure memory architecture.

The Problem

Traditional approach: Load every SKILL.md to decide which skill to use.

• 30+ skills × ~3000 tokens each = 90,000+ tokens just for discovery
• Agent only needs ~500 tokens to make the choice
• 89,500 tokens wasted on reading full skills

The Solution: 3-Layer Loading

┌────────────────────────────────────────────────┐
│ LAYER 1: INDEX (~100 tokens per skill)          │
│ Always loaded. Quick scan. "What exists?"        │
│ Name + Domain + Triggers + 1-line summary       │
└─────────────────────┬──────────────────────────┘
                      │ (match found?)
                      v
┌────────────────────────────────────────────────┐
│ LAYER 2: SUMMARY (~300 tokens per skill)        │
│ Load when choosing. "Is this the right one?"     │
│ Description + When to use + Integration table   │
└─────────────────────┬──────────────────────────┘
                      │ (confirmed?)
                      v
┌────────────────────────────────────────────────┐
│ LAYER 3: FULL SKILL.md (1500-5000 tokens)       │
│ Load ONLY during execution. "How to use it."     │
│ Complete instructions, examples, red flags       │
└────────────────────────────────────────────────┘

Layer 1: Skill Index

Always loaded by agents at session start. Maximum 100 tokens per skill.

Engineering Swarm 🔧

cm-tdd

cm-debugging

cm-quality-gate

cm-test-gate

cm-code-review

cm-codeintell

cm-reactor

cm-clean-code

Operations Swarm ⚙️

cm-safe-deploy

cm-identity-guard

cm-git-worktrees

cm-terminal

Security Swarm 🔒

cm-secret-shield

Product Swarm 🎨

cm-brainstorm-idea

cm-planning

cm-ux-master

cm-ui-preview

cm-dockit

cm-project-bootstrap

cm-jtbd

Growth Swarm 📈

cm-content-factory

cm-ads-tracker

cm-cro-methodology

Orchestration Swarm 🎯

cm-execution

cm-continuity

cm-skill-mastery

cm-skill-health

cm-skill-evolution

cm-skill-search

cm-skill-share

cm-safe-i18n

cm-skill-chain

cm-deep-search

cm-readit

cm-how-it-work

Workflow Commands 🖥️

cm-start

cm-dashboard

tasks.md

cm-tasks.json

cm-status

Planned (not yet implemented) 🧪

Skills below do not have SKILL.md files yet.

mom-test

release-it

medical-research

tailwind-mastery

pandasai-analytics

cm-google-form

Community Skills Registry 🌐

STOP! Zero-Token Discovery Priority: BEFORE using this registry, check if

.cm/project-skills.md

exists in the local project. It contains the pre-compiled, exact community skills needed for the project's tech stack. If it doesn't exist, run

npx cm index skills

first to build it.

Only use the registry below if tech is not listed or if you need to search manually:

npx skills find "{keyword}"

🖥️ Frontend Frameworks (10)

react-best-practices

composition-patterns

react-native-guidelines

react-native-best-practices

nextjs-app-router-patterns

react:components

cm-ux-master

angular-directives

svelte-best-practices

astro-best-practices

🎨 CSS & UI Design (8)

web-design-guidelines

tailwind-design-system

tailwind-v4-shadcn

tailwindcss-advanced-layouts

remotion-best-practices

shadcn-ui-patterns

css-architecture

framer-motion

🗄️ Databases & ORM (9)

supabase-postgres-best-practices

prisma-database-setup

prisma-client-api

prisma-expert

drizzle-orm

mongodb-patterns

pinecone-db

turso-database

planetscale

🔐 Authentication (6)

clerk-nextjs-patterns

clerk-auth-v5

better-auth-best-practices

next-auth-v5

auth-js-v5

supabase-auth

💳 Payments & Commerce (4)

stripe-client

stripe-webhook

stripe-checkout

stripe-connect

🧪 Testing (5)

testing-library

vitest-best-practices

playwright-e2e

cypress-patterns

testing-guidelines

☁️ Cloud & Deployment (7)

vercel-deploy-claimable

vercel-deployment-best-practices

azure-deployment

aws-deployment

docker-expert

docker-compose-orchestration

multi-stage-dockerfile

🤖 AI & ML (5)

vercel-ai-sdk-best-practices

langchain-skills

openai-skills

tavily-search

ai-sdk-patterns

📡 APIs & Integrations (3)

nextjs-supabase-auth

supabase-edge-functions

nextjs-seo

📱 Mobile & Cross-Platform (2)

expo-patterns

capacitor-patterns

Discovery Protocol:

1. Check THIS index first (59 skills above)
   → Match found? → npx skills add {source} --skill {name} -a antigravity
2. No match? → npx skills find "{keyword}" on skills.sh
3. Still nothing? → Proceed without skill, note gap in .cm-skills-log.json

Trigger points:
  - cm-start Step 0.5 (before planning)
  - cm-planning Phase A Step 4 (scope analysis)
  - cm-execution Pre-flight (before dispatch)
  - cm-execution RARV (after 2 failed retries)

Layer 2: Skill Summary

Loaded when an agent has matched a skill from Layer 1 and needs more context.

To access Layer 2, read the first 20 lines of the corresponding SKILL.md (frontmatter + first section).

Example: cm-tdd Layer 2
  - Full name: "TDD — Test Before Write"
  - When to Use: ALWAYS before implementing any feature or bugfix
  - Process: Red (write failing test) → Green (minimal code) → Refactor (clean up)
  - Integration: Works with cm-execution (subagent TDD), cm-quality-gate (coverage check)
  - Token cost: ~4200 tokens for full skill

Layer 3: Full SKILL.md

Loaded ONLY when executing. Read the complete SKILL.md file.

# Agent loads full skill only when ready to execute
view_file /path/to/skills/cm-tdd/SKILL.md

Usage Protocol

1. AT SESSION START:
   - Load this skill index (Layer 1) — costs ~2500 tokens total
   - Now you know what 34 skills do without reading any of them

2. WHEN MATCHING A TASK:
   - Scan Layer 1 triggers to find matching skill(s)
   - If unsure, read Layer 2 (first 20 lines of SKILL.md)
   - Choose the best skill

3. WHEN EXECUTING:
   - Load Layer 3 (full SKILL.md) for the chosen skill
   - Follow the skill's instructions completely

4. SAVINGS:
    - Without index: 34 × 3000 = 102,000 tokens
    - With index: 2800 + 300 + 3000 = 6,100 tokens
    - Saved: ~95,900 tokens (~94%)

Skill Domain Mapping

For dynamic agent selection (Phase 3), skills map to domains:

engineering: [cm-tdd, cm-debugging, cm-quality-gate, cm-test-gate, cm-code-review, cm-codeintell, cm-reactor, cm-clean-code]
operations: [cm-safe-deploy, cm-identity-guard, cm-git-worktrees, cm-terminal, cm-safe-i18n]
security: [cm-secret-shield]
product: [cm-brainstorm-idea, cm-planning, cm-ux-master, cm-ui-preview, cm-dockit, cm-project-bootstrap, cm-jtbd]
growth: [cm-content-factory, cm-ads-tracker, cm-cro-methodology]
orchestration: [cm-execution, cm-continuity, cm-skill-mastery, cm-skill-health, cm-skill-evolution, cm-skill-search, cm-skill-share, cm-skill-chain, cm-deep-search, cm-readit, cm-how-it-work]
commands: [cm-start, cm-dashboard, cm-status]
# planned (no SKILL.md yet): mom-test, release-it, medical-research, tailwind-mastery, pandasai-analytics, cm-google-form

The Bottom Line

Know everything. Load nothing. Execute precisely. Save 90%+ tokens.