DollhouseMCP

# Unified Capability Index for DollhouseMCP ## Optimized LLM Priming Strategy Created: September 22, 2025 Based on: Empirical testing results from capability index tests ## Critical Design Principles ### 1. Verb-Based Action Triggers (Not Nouns) - Use action words that match user intent - Examples: debug, fix, explain, analyze, create, search, find - NOT: debugger, fixer, explainer, analyst ### 2. Flat Structure with High Attention Surface - Simple key-value pairs at top level - No deep nesting that reduces attention probability - Direct mappings visible immediately ### 3. Progressive Disclosure Pattern - Level 1: Triggers (10 tokens) - Always visible - Level 2: Summaries (50 tokens) - Load on match - Level 3: Full content (150-500 tokens) - Load on demand ## The Unified Index Structure ```yaml # DOLLHOUSE_CAPABILITY_INDEX # This index MUST appear at the TOP of any CLAUDE.md or context CAPABILITY_TRIGGERS: # Debugging & Errors debug → debug-detective error → debug-detective crash → debug-detective fix → debug-detective broken → debug-detective # Explanation & Learning explain → eli5-explainer teach → eli5-explainer simplify → eli5-explainer understand → eli5-explainer # Security & Analysis security → security-analyst vulnerability → security-analyst audit → security-analyst analyze → technical-analyst investigate → technical-analyst # Creation & Writing write → creative-writer story → creative-writer create → creative-writer imagine → creative-writer # Business & Strategy roi → business-consultant strategy → business-consultant business → business-consultant profit → business-consultant # Memory Operations remember → memory-manager recall → memory-manager forgot → memory-manager previous → memory-manager session → memory-manager # Git & Version Control git → git-workflow commit → git-workflow branch → git-workflow merge → git-workflow # Docker & Containers docker → docker-memory container → docker-memory authentication → docker-memory # GitHub Operations issue → github-tools pr → github-tools pull-request → github-tools # Testing & Quality test → test-runner verify → test-runner validate → test-runner check → test-runner # CAPABILITY_SUMMARIES (Load when trigger matches) debug-detective: DO: "Systematic debugging, root cause analysis, error isolation" USE_WHEN: "error|bug|crash|failure|broken|not working|undefined" ACTION: "activate_element('debug-detective', 'personas')" MEMORIES: ["javascript-errors", "python-errors", "common-bugs"] TOKEN_COST: 145 eli5-explainer: DO: "Simple explanations using analogies and examples" USE_WHEN: "explain|teach|simplify|understand|confusing|complex" ACTION: "activate_element('eli5-explainer', 'personas')" MEMORIES: ["teaching-techniques", "common-analogies"] TOKEN_COST: 132 security-analyst: DO: "Security vulnerability detection and remediation" USE_WHEN: "security|vulnerability|exploit|injection|xss|csrf" ACTION: "activate_element('security-analyst', 'personas')" MEMORIES: ["owasp-top-10", "security-best-practices"] TOKEN_COST: 156 memory-manager: DO: "Retrieve and manage persistent memories" USE_WHEN: "remember|recall|forgot|previous|earlier|session" SEARCHES: - "search_portfolio(query)" - "list_elements('memories')" - "get_element_details(name, 'memories')" TOKEN_COST: 98 git-workflow: DO: "Git operations and version control" USE_WHEN: "git|commit|branch|merge|rebase|cherry-pick" MEMORIES: ["gitflow-strategy", "commit-conventions"] TOOLS: ["gh", "git"] TOKEN_COST: 112 docker-memory: CONTAINS: "Docker authentication solution for Claude Code" USE_WHEN: "docker|container|authentication|apiKeyHelper" ACTION: "get_element_details('docker-claude-code-authentication-solution', 'memories')" TOKEN_COST: 287 github-tools: DO: "GitHub issue and PR management" USE_WHEN: "issue|pr|pull request|review|merge" TOOLS: ["gh pr", "gh issue"] TEMPLATES: ["pr-template", "issue-template"] TOKEN_COST: 134 # MEMORY_SELECTION_HINTS (For retrieval scenarios) MEMORY_PRIORITIES: # When multiple memories match, prefer by recency and specificity javascript-errors: CONTAINS: "TypeError, undefined properties, optional chaining" PRIORITY: 1 # Highest for JS errors session-recent: PATTERN: "session-2025-09-*" PRIORITY: 2 # Recent sessions second generic-debugging: CONTAINS: "General debugging strategies" PRIORITY: 5 # Lowest priority # TOOL_SELECTION_RULES TOOL_PREFERENCES: # Direct mappings for common operations list_any: "list_elements(type)" search_local: "search_portfolio(query)" search_community: "search_collection(query)" get_details: "get_element_details(name, type)" activate: "activate_element(name, type)" deactivate: "deactivate_element(name, type)" # ACTIVATION_PATTERNS AUTO_ACTIVATE: # Patterns that should trigger immediate activation "fix this error": - activate_element('debug-detective', 'personas') - search_portfolio('error') "explain simply": - activate_element('eli5-explainer', 'personas') "check security": - activate_element('security-analyst', 'personas') - search_portfolio('security') "what did we do": - search_portfolio('session') - list_elements('memories') ``` ## Implementation Strategy ### 1. Where to Place the Index The index should be injected at the **TOP** of context in this order: 1. System prompt 2. **CAPABILITY_INDEX** (this document) 3. User's CLAUDE.md 4. Other context ### 2. Token Budget ``` Always Loaded (Level 1): - Trigger map: ~50 tokens - Tool preferences: ~30 tokens - Total: ~80 tokens Conditionally Loaded (Level 2): - Relevant summaries: ~50 tokens each - Usually 2-3 match: ~100-150 tokens On-Demand (Level 3): - Full capability: ~150-500 tokens - Only when executing Total Average: ~250-350 tokens vs 7,000+ without index ``` ### 3. Validation Testing To verify the index works: 1. **Trigger Test**: Use each trigger word, verify correct capability loads 2. **Memory Selection**: Create 3 similar memories, verify correct selection 3. **Cascade Test**: Monitor token usage at each level 4. **Heavy Context**: Embed in 25k context, measure performance ### 4. Expected Behaviors When working correctly, you should see: ``` User: "I need to debug this error" AI: [Sees "debug" → loads debug-detective summary] [Activates debug-detective persona] [Searches for relevant error memories] [Provides targeted solution] Tokens used: ~400 (vs ~2,000 without index) ``` ### 5. Continuous Improvement The index should be updated based on: - Most frequently used capabilities - Failed activation attempts - New elements added to portfolio - User feedback on response quality ## Metrics for Success 1. **Token Reduction**: Target 80% reduction in memory token usage 2. **Activation Accuracy**: >90% correct capability selection 3. **Response Time**: <100ms to select capability 4. **Memory Selection**: >80% correct memory retrieval 5. **User Satisfaction**: Fewer manual activations needed ## Next Steps 1. Implement index generation from existing elements 2. Add to MCP server startup routine 3. Inject at top of all contexts 4. Monitor and measure performance 5. Iterate based on real usage data --- *This unified index represents the optimal priming strategy based on empirical testing and token mechanics analysis.*