@agile-vibe-coding/avc 0.1.0 → 0.2.3

package/cli/evaluation-prompts.js

@@ -0,0 +1,1008 @@
+/**
+ * Evaluation prompts for LLM model selection across all AVC ceremonies and stages
+ * Extracted from AVC_DEFAULT_LLMS.md
+ *
+ * Each prompt is used to query LLM providers about which of their models
+ * is best suited for the specific stage's requirements.
+ */
+
+export const EVALUATION_PROMPTS = [
+  {
+    id: 'sponsor-call-suggestions',
+    ceremony: 'sponsor-call',
+    stage: 'suggestions',
+    stageName: 'Questionnaire Suggestions',
+    prompt: `TASK: Select optimal LLM model for questionnaire suggestions in sponsor-call ceremony
+PRIORITY: Best possible output quality
+
+CONTEXT:
+- Stage: suggestions (Questionnaire Suggestions)
+- Ceremony: sponsor-call
+- Purpose: AI analyzes project name and suggests intelligent, contextually appropriate answers for 6 questionnaire fields
+- Input: Project name only (10-50 tokens)
+- Output: Suggested answers for Mission Statement, Target Users, Initial Scope, Deployment Target, Technical Considerations, Security Requirements (500-1500 tokens total)
+- Call frequency: 1 per ceremony
+- User interaction: Real-time (user waiting for suggestions while viewing questionnaire)
+- Impact: Sets initial project direction and quality tone for all downstream artifacts
+
+EVALUATION CRITERIA (BEST OUTPUT FIRST):
+
+1. Output Quality Requirements: HIGH
+   - Must infer project domain from minimal context (just project name)
+   - Suggestions must be relevant, specific, and actionable
+   - Must demonstrate domain knowledge and industry best practices
+   - Quality directly impacts user's project definition experience
+   → Requires appropriate model capabilities
+
+2. Task Complexity: 6/10 (Moderate reasoning with domain inference)
+   - Analyze project name for domain clues
+   - Infer appropriate technology stack and architecture
+   - Generate contextually relevant suggestions across 6 different categories
+   - Balance between generic and specific recommendations
+
+3. Context Understanding:
+   - Extremely limited input context (just project name)
+   - Must leverage broad domain knowledge to compensate
+   - Must infer user intent and project type
+   → Requires appropriate model capabilities
+
+4. Consistency & Reliability:
+   - Suggestions must be coherent across all 6 fields
+   - Must avoid contradictory recommendations
+   - Critical for user's first impression of AVC
+   → Requires appropriate model capabilities
+
+5. Speed Requirements: IMPORTANT (Secondary)
+   - User is actively waiting for suggestions
+   - Real-time interaction requires reasonable response time
+   - 2-4 second response ideal, <8 seconds acceptable
+
+6. Pricing Considerations: TERTIARY
+   - Single call per ceremony (low volume)
+   - User-facing quality important
+   - Pricing impact minimal with only 1 call
+   → Quality worth any pricing tier for this stage
+
+RECOMMENDATION:
+
+Based on evaluation criteria above, select the model that best meets:
+1. PRIMARY: Output Quality - Excellent domain inference from minimal context (project name)
+2. SECONDARY: World Knowledge - Strong understanding of business domains and technical stacks
+3. TERTIARY: Task Complexity - Moderate (7/10) - requires intelligent suggestion generation
+4. Speed Requirements: MODERATE - User waiting in real-time (2-4s ideal, <8s acceptable)
+5. Pricing: TERTIARY - Single call per ceremony, minimal impact
+
+Selection Guidance:
+- Prioritize models with strong domain/business knowledge for accurate suggestions
+- Require excellent inference capabilities from very limited context
+- Must support real-time interaction speed (<8 seconds)
+- Pricing is not a constraint (quality worth any tier for user-facing interaction)
+
+Consider all available models from any provider (Claude, Gemini, OpenAI, etc.)`,
+    metadata: {
+      callFrequency: 1,
+      impact: 'HIGH',
+      taskComplexity: 6,
+      currentDefault: 'Claude Sonnet 4.5'
+    }
+  },
+
+  {
+    id: 'sponsor-call-documentation',
+    ceremony: 'sponsor-call',
+    stage: 'documentation',
+    stageName: 'Documentation Generation',
+    prompt: `TASK: Select optimal LLM model for documentation generation in sponsor-call ceremony
+PRIORITY: Best possible output quality
+
+CONTEXT:
+- Stage: documentation (Project Documentation Creation)
+- Ceremony: sponsor-call
+- Purpose: AI generates comprehensive PROJECT.md from questionnaire answers
+- Input: 6 questionnaire field answers (1,000-5,000 tokens)
+- Output: Structured PROJECT.md with Executive Summary, Problem Statement, Solution, User Personas, Core Features, Technical Architecture, Security, Success Metrics (2,000-8,000 tokens)
+- Call frequency: 1 per ceremony
+- User interaction: Background process after questionnaire completion
+- Impact: CRITICAL - This is the foundational document that defines the entire project; all future AI agents, epics, stories, and tasks derive from this document
+
+EVALUATION CRITERIA (BEST OUTPUT FIRST):
+
+1. Output Quality Requirements: CRITICAL
+   - Must produce well-structured, comprehensive technical documentation
+   - Content must be coherent, professional, and actionable
+   - Must maintain consistency across all sections
+   - Quality affects all downstream project artifacts
+   → Requires appropriate model capabilities
+
+2. Task Complexity: 8/10 (Advanced technical writing and synthesis)
+   - Synthesize multiple questionnaire inputs into coherent narrative
+   - Organize information into logical section structure
+   - Generate appropriate technical architecture descriptions
+   - Create realistic user personas and success metrics
+   - Maintain professional tone and technical accuracy
+
+3. Context Understanding:
+   - Must understand relationships between questionnaire answers
+   - Infer appropriate technical depth and detail level
+   - Expand brief answers into comprehensive sections
+   - Maintain consistency across document sections
+   → Requires appropriate model capabilities
+
+4. Consistency & Reliability:
+   - Critical that all sections align and don't contradict
+   - Technical architecture must match scope and requirements
+   - Success metrics must align with stated goals
+
+5. Speed Requirements: LOW (Secondary)
+   - Background process, user not actively waiting
+   - Quality far more important than speed
+   - Can take 10-30 seconds without issue
+   → Speed not a constraint
+
+6. Pricing Considerations: TERTIARY
+   - Single call per ceremony
+   - Foundational document for entire project
+   - Quality impact is massive
+   → Worth any pricing tier given criticality
+
+RECOMMENDATION:
+
+Based on evaluation criteria above, select the model that best meets:
+1. PRIMARY: Output Quality - CRITICAL - Foundational project document
+2. SECONDARY: Technical Writing - Excellent multi-section document structure
+3. TERTIARY: Task Complexity - High (8/10) - requires synthesis of diverse inputs
+4. Speed Requirements: LOW - Background process, quality >>> speed
+5. Pricing: TERTIARY - Single call, massive quality impact justifies any tier
+
+Selection Guidance:
+- Prioritize models with exceptional technical writing and document synthesis
+- Require strong markdown formatting and organizational structure
+- Must maintain consistency across complex 9-section document
+- Must synthesize questionnaire inputs, suggestions, and context into coherent narrative
+- Speed is not a constraint (can take 10-30 seconds)
+- Pricing not a limitation (worth premium for project foundation)
+
+Consider all available models from any provider (Claude, Gemini, OpenAI, etc.)`,
+    metadata: {
+      callFrequency: 1,
+      impact: 'CRITICAL',
+      taskComplexity: 8,
+      currentDefault: 'Claude Sonnet 4.5'
+    }
+  },
+
+  {
+    id: 'sponsor-call-validation',
+    ceremony: 'sponsor-call',
+    stage: 'validation',
+    stageName: 'Documentation Validation',
+    prompt: `TASK: Select optimal LLM model for documentation validation in sponsor-call ceremony
+PRIORITY: Best possible output quality
+
+CONTEXT:
+- Stage: validation (Documentation & Context Validation)
+- Ceremony: sponsor-call
+- Purpose: AI validators check PROJECT.md and context.md against quality rules
+- Input: Full PROJECT.md or context.md + validation rules (5,000-12,000 tokens)
+- Output: Validation report identifying issues and suggestions
+- Call frequency: 2 validators per ceremony
+- Impact: HIGH - Catches quality issues before they propagate
+
+EVALUATION CRITERIA (BEST OUTPUT FIRST):
+
+1. Output Quality Requirements: HIGH
+   - Must accurately identify quality issues and inconsistencies
+   - Must provide actionable feedback for improvement
+   - Must understand project documentation best practices
+   → Requires appropriate model capabilities
+
+2. Task Complexity: 6/10 (Analytical validation + rule application)
+   - Apply validation rules to documentation
+   - Identify inconsistencies and gaps
+   - Assess completeness and coherence
+
+3. Context Understanding:
+   - Must handle large documents (full PROJECT.md)
+   - Understand relationships across sections
+   - Identify subtle inconsistencies
+   → Requires appropriate model capabilities
+
+4. Speed Requirements: LOW (Secondary)
+   - Background validation stage
+   - Quality more important than speed
+
+6. Pricing Considerations: TERTIARY
+   - Only 2 calls per ceremony
+   - Moderate quality requirements
+
+RECOMMENDATION:
+
+Based on evaluation criteria above, select the model that best meets:
+1. PRIMARY: Output Quality - HIGH - Analytical validation with actionable feedback
+2. SECONDARY: Task Complexity - Moderate (6/10) - rule application + gap identification
+3. TERTIARY: Context Understanding - Must handle large documents with cross-section relationships
+4. Speed Requirements: LOW - Background validation, quality >>> speed
+5. Pricing: TERTIARY - 2 calls per ceremony, moderate impact
+
+Selection Guidance:
+- Prioritize models with strong analytical and critical thinking capabilities
+- Require ability to identify subtle inconsistencies and gaps
+- Must provide actionable, specific feedback for improvement
+- Must understand project documentation best practices
+- Speed not a constraint (background processing)
+- Consider pricing tier relative to quality improvement
+
+Consider all available models from any provider (Claude, Gemini, OpenAI, etc.)`,
+    metadata: {
+      callFrequency: 2,
+      impact: 'HIGH',
+      taskComplexity: 6,
+      currentDefault: 'Claude Sonnet 4.5'
+    }
+  },
+
+  {
+    id: 'sprint-planning-decomposition',
+    ceremony: 'sprint-planning',
+    stage: 'decomposition',
+    stageName: 'Epic & Story Decomposition',
+    prompt: `TASK: Select optimal LLM model for epic and story decomposition in sprint-planning ceremony
+PRIORITY: Best possible output quality
+
+CONTEXT:
+- Stage: decomposition (Epic & Story Decomposition)
+- Ceremony: sprint-planning
+- Purpose: AI analyzes PROJECT.md and decomposes project scope into hierarchical epics and stories
+- Input: PROJECT.md, project context.md, existing epics/stories (5,000-20,000 tokens)
+- Output: Structured JSON with epics, stories, acceptance criteria
+- Call frequency: 1 per ceremony
+- Impact: CRITICAL - Defines entire project work breakdown structure
+
+EVALUATION CRITERIA (BEST OUTPUT FIRST):
+
+1. Output Quality Requirements: CRITICAL
+   - Must create logical, implementable work breakdown
+   - Acceptance criteria must be specific and testable
+   - Hierarchy must reflect dependencies
+   → Requires appropriate model capabilities
+
+2. Task Complexity: 9/10 (Complex hierarchical decomposition)
+   - Analyze full project scope
+   - Break features into implementable story-level units
+   - Generate specific, testable acceptance criteria
+   - Identify dependencies between stories
+
+3. Context Understanding:
+   - Must synthesize entire project vision
+   - Understand technical architecture and constraints
+   - Recognize implicit dependencies
+   → Requires appropriate model capabilities
+
+4. Consistency & Reliability:
+   - Critical that decomposition is complete (no gaps)
+   - Stories must not overlap or contradict
+
+5. Speed Requirements: LOW (Secondary)
+   - Background process, one-time operation
+   - Quality far more important than speed
+
+6. Pricing Considerations: TERTIARY
+   - Single call per ceremony
+   - Most critical stage in sprint-planning
+   → Worth any pricing tier
+
+RECOMMENDATION:
+
+Based on evaluation criteria above, select the model that best meets:
+1. PRIMARY: Output Quality - CRITICAL - Defines entire project work structure
+2. SECONDARY: Task Complexity - Very High (9/10) - complex hierarchical decomposition
+3. TERTIARY: Hierarchical Reasoning - Exceptional domain breakdown with dependencies
+4. Speed Requirements: LOW - Background process, quality >>> speed
+5. Pricing: TERTIARY - Single call, most critical stage, worth any tier
+
+Selection Guidance:
+- Prioritize models with exceptional hierarchical reasoning and decomposition
+- Require deep project synthesis and implicit dependency recognition
+- Must generate complete, non-overlapping, consistent epic/story structure
+- Must produce specific, testable acceptance criteria at story level
+- Consistency and completeness are paramount (gaps cause downstream failures)
+- Speed not a constraint (one-time background operation)
+- Pricing justified by critical impact on entire project
+
+Consider all available models from any provider (Claude, Gemini, OpenAI, etc.)`,
+    metadata: {
+      callFrequency: 1,
+      impact: 'CRITICAL',
+      taskComplexity: 9,
+      currentDefault: 'Claude Opus 4.6'
+    }
+  },
+
+  {
+    id: 'sprint-planning-validation-universal',
+    ceremony: 'sprint-planning',
+    stage: 'validation-universal',
+    stageName: 'Universal Validators',
+    prompt: `TASK: Select optimal LLM model for universal validators in sprint-planning ceremony
+PRIORITY: Best possible output quality
+
+CONTEXT:
+- Stage: validation-universal (Universal Validators)
+- Ceremony: sprint-planning
+- Purpose: Critical validators always applied: architecture, security, quality
+- Input: Epic or Story with full context (2,000-8,000 tokens per item)
+- Output: Detailed validation report with architectural and security analysis
+- Call frequency: ~30 calls per ceremony
+- Impact: CRITICAL - Catches fundamental issues before implementation
+
+EVALUATION CRITERIA (BEST OUTPUT FIRST):
+
+1. Output Quality Requirements: CRITICAL
+   - Must perform deep architectural analysis
+   - Must identify security vulnerabilities
+   - Must provide specific, actionable recommendations
+   → Requires appropriate model capabilities
+
+2. Task Complexity: 9/10 (Deep architectural and security analysis)
+   - Analyze system architecture and design patterns
+   - Identify security vulnerabilities
+   - Assess technical feasibility
+   - Evaluate testing strategies
+
+3. Context Understanding:
+   - Must understand full project architecture
+   - Cross-reference with other epics/stories
+   - Identify system-wide architectural issues
+   → Requires appropriate model capabilities
+
+4. Consistency & Reliability:
+   - Cannot miss critical architectural flaws
+   - Must consistently identify security issues
+   → Critical appropriate model capabilities
+
+5. Speed Requirements: MODERATE (Secondary)
+   - 30 calls in parallel validation stage
+   - Quality far more important than speed
+
+6. Pricing Considerations: MODERATE
+   - 30 calls = significant volume
+   - Quality cannot be compromised
+
+RECOMMENDATION:
+
+Based on evaluation criteria above, select the model that best meets:
+1. PRIMARY: Output Quality - CRITICAL - Deep architectural and security analysis
+2. SECONDARY: Task Complexity - Very High (9/10) - system-wide analysis
+3. TERTIARY: Reliability - Cannot miss critical architectural flaws or security issues
+4. Speed Requirements: MODERATE - 30 calls in parallel, quality still paramount
+5. Pricing: MODERATE - 30 calls = significant volume, but quality cannot be compromised
+
+Selection Guidance:
+- Prioritize models with exceptional architectural reasoning and security analysis
+- Require consistent, reliable identification of vulnerabilities and design flaws
+- Must provide specific, actionable recommendations for improvements
+- Must understand system-wide architecture and cross-reference between components
+- Reliability is critical (false negatives in security/architecture are high-risk)
+- Do not compromise on quality for these critical validators
+
+Consider all available models from any provider (Claude, Gemini, OpenAI, etc.)`,
+    metadata: {
+      callFrequency: 30,
+      impact: 'CRITICAL',
+      taskComplexity: 9,
+      currentDefault: 'Claude Sonnet 4.5'
+    }
+  },
+
+  {
+    id: 'sprint-planning-validation-domain',
+    ceremony: 'sprint-planning',
+    stage: 'validation-domain',
+    stageName: 'Domain Validators',
+    prompt: `TASK: Select optimal LLM model for domain validators in sprint-planning ceremony
+PRIORITY: Best possible output quality
+
+CONTEXT:
+- Stage: validation-domain (Domain Validators)
+- Ceremony: sprint-planning
+- Purpose: Domain-specific validators based on tech stack
+- Input: Epic or Story with domain-relevant context (2,000-5,000 tokens)
+- Output: Domain-specific validation report with best practices
+- Call frequency: ~90 calls per ceremony (largest volume stage)
+- Impact: HIGH - Catches domain-specific issues, ensures best practices
+
+EVALUATION CRITERIA (BEST OUTPUT FIRST):
+
+1. Output Quality Requirements: HIGH
+   - Must apply domain-specific best practices
+   - Must identify domain anti-patterns
+   - Must provide actionable recommendations
+   → Requires appropriate model capabilities
+
+2. Task Complexity: 7/10 (Domain expertise application)
+   - Apply domain-specific patterns
+   - Identify violations of domain conventions
+   - Not just pattern matching - requires context understanding
+
+3. Context Understanding:
+   - Must understand project architecture in domain context
+   - Cross-reference with other domains
+
+4. Speed Requirements: MODERATE (Secondary)
+   - 90 calls = highest volume in sprint-planning
+   - Parallel execution
+
+6. Pricing Considerations: SIGNIFICANT
+   - 90 calls = largest pricing driver
+   - Medium vs Low pricing makes material difference
+
+RECOMMENDATION:
+
+Based on evaluation criteria above, select the model that best meets:
+1. PRIMARY: Output Quality - HIGH - Domain-specific best practices and anti-patterns
+2. SECONDARY: Task Complexity - High (7/10) - requires domain expertise, not just pattern matching
+3. TERTIARY: Volume Impact - HIGHEST (90 calls) - largest pricing driver in ceremony
+4. Speed Requirements: MODERATE - Parallel execution, but quality important
+5. Pricing: SIGNIFICANT - 90 calls make pricing tier materially impactful
+
+Selection Guidance:
+- Prioritize models with strong domain knowledge (DevOps, Database, API, Frontend, etc.)
+- Require ability to apply domain-specific best practices and identify anti-patterns
+- Must provide actionable, context-appropriate recommendations
+- Consider pricing tier carefully (90 calls = 3-4x cost difference between tiers)
+- Balance domain expertise depth vs pricing efficiency
+- Acceptable to use lower tier if basic domain pattern checking sufficient
+
+Consider all available models from any provider (Claude, Gemini, OpenAI, etc.)`,
+    metadata: {
+      callFrequency: 90,
+      impact: 'HIGH',
+      taskComplexity: 7,
+      currentDefault: 'Claude Sonnet 4.5'
+    }
+  },
+
+  {
+    id: 'sprint-planning-validation-feature',
+    ceremony: 'sprint-planning',
+    stage: 'validation-feature',
+    stageName: 'Feature Validators',
+    prompt: `TASK: Select optimal LLM model for feature validators in sprint-planning ceremony
+PRIORITY: Best possible output quality
+
+CONTEXT:
+- Stage: validation-feature (Feature Validators)
+- Ceremony: sprint-planning
+- Purpose: Feature-specific validators based on keywords in acceptance criteria
+- Input: Epic or Story with feature-specific context (1,500-4,000 tokens)
+- Output: Feature-specific validation checklist and completeness assessment
+- Call frequency: ~25 calls per ceremony
+- Impact: MEDIUM - Ensures feature completeness, identifies missing requirements
+
+EVALUATION CRITERIA (BEST OUTPUT FIRST):
+
+1. Output Quality Requirements: MEDIUM-HIGH
+   - Must verify feature completeness
+   - Must identify missing scenarios or edge cases
+   - Should provide feature-specific implementation guidance
+   → Requires appropriate model capabilities
+
+2. Task Complexity: 5/10 (Feature checklist validation)
+   - Apply feature-specific checklists
+   - Identify missing scenarios
+   - Verify acceptance criteria completeness
+
+3. Context Understanding:
+   - Understand feature requirements from acceptance criteria
+   - Identify implicit requirements not explicitly stated
+
+4. Speed Requirements: MODERATE (Secondary)
+   - 25 calls, parallel execution
+   - Background processing
+
+6. Pricing Considerations: MODERATE
+   - 25 calls = moderate volume
+   - Completeness checking vs deep analysis
+
+RECOMMENDATION:
+
+Based on evaluation criteria above, select the model that best meets:
+1. PRIMARY: Output Quality - MEDIUM-HIGH - Feature completeness and edge case identification
+2. SECONDARY: Task Complexity - Moderate (5/10) - checklist application with implicit requirement detection
+3. TERTIARY: Volume Impact - MODERATE (25 calls) - pricing tier makes moderate impact
+4. Speed Requirements: MODERATE - Parallel execution in background
+5. Pricing: MODERATE - 25 calls = moderate volume, balance quality vs cost
+
+Selection Guidance:
+- Prioritize models capable of identifying implicit requirements (not just explicit ones)
+- Require feature completeness checking (missing scenarios, edge cases)
+- Should provide test scenario suggestions
+- Consider pricing tier (25 calls = 2-3x cost difference between tiers)
+- Balance edge case detection quality vs pricing efficiency
+- Acceptable to use lower tier if basic explicit requirement checking sufficient
+
+Consider all available models from any provider (Claude, Gemini, OpenAI, etc.)`,
+    metadata: {
+      callFrequency: 25,
+      impact: 'MEDIUM',
+      taskComplexity: 5,
+      currentDefault: 'Claude Sonnet 4.5'
+    }
+  },
+
+  {
+    id: 'sprint-planning-doc-distribution',
+    ceremony: 'sprint-planning',
+    stage: 'doc-distribution',
+    stageName: 'Documentation Distribution',
+    prompt: `TASK: Select optimal LLM model for documentation distribution in sprint-planning ceremony
+PRIORITY: Best possible output quality
+
+CONTEXT:
+- Stage: doc-distribution (Documentation Distribution)
+- Ceremony: sprint-planning
+- Purpose: AI moves relevant content from parent doc.md to each epic/story doc.md and elaborates with domain-specific detail
+- Input: Parent doc.md + epic/story item description (2,000-5,000 tokens)
+- Output: Child doc.md (extracted + elaborated) + updated parent doc.md (lighter) as JSON
+- Call frequency: ~25 calls per ceremony
+- Impact: VERY HIGH - Establishes the documentation tree that guides all future AI agents
+
+EVALUATION CRITERIA (BEST OUTPUT FIRST):
+
+1. Output Quality Requirements: VERY HIGH
+   - Must accurately identify which content belongs to the child scope
+   - Must elaborate with domain-specific implementation detail
+   - Must produce valid JSON with embedded markdown
+   → Requires appropriate model capabilities
+
+2. Task Complexity: 7/10 (Content extraction + elaboration + JSON output)
+   - Identify domain-relevant sections in parent doc
+   - Remove extracted content cleanly from parent
+   - Extend child doc with actionable implementation notes
+
+3. Consistency & Reliability:
+   - Child doc must cover only the child's scope
+   - Parent doc must remain coherent after extraction
+   - JSON output must be well-formed
+
+4. Speed Requirements: MODERATE (Secondary)
+   - 25 calls, sequential per epic/story chain
+   - Quality more important than speed
+
+5. Pricing Considerations: MODERATE
+   - 25 calls = moderate volume
+   - High impact on downstream work quality
+
+RECOMMENDATION:
+
+Based on evaluation criteria above, select the model that best meets:
+1. PRIMARY: Output Quality - VERY HIGH - Defines the documentation hierarchy for the project
+2. SECONDARY: Technical Writing - Focused extraction with domain elaboration
+3. TERTIARY: JSON reliability - Must produce parseable JSON with embedded markdown
+
+Consider all available models from any provider (Claude, Gemini, OpenAI, etc.)`,
+    metadata: {
+      callFrequency: 25,
+      impact: 'VERY HIGH',
+      taskComplexity: 7,
+      currentDefault: 'Claude Sonnet 4.6'
+    }
+  },
+
+  {
+    id: 'seed-decomposition',
+    ceremony: 'seed',
+    stage: 'decomposition',
+    stageName: 'Task Decomposition',
+    prompt: `TASK: Select optimal LLM model for task decomposition in seed ceremony
+PRIORITY: Best possible output quality
+
+CONTEXT:
+- Stage: decomposition (Task & Subtask Decomposition)
+- Ceremony: seed
+- Purpose: AI breaks down story into tasks and subtasks
+- Input: Story with acceptance criteria, contexts (2,000-6,000 tokens)
+- Output: Hierarchical task structure with dependencies
+- Call frequency: 1 per ceremony execution
+- Impact: CRITICAL - Defines actual implementation plan
+
+EVALUATION CRITERIA (BEST OUTPUT FIRST):
+
+1. Output Quality Requirements: CRITICAL
+   - Must break story into logical task groupings
+   - Subtasks must be atomic and implementable
+   - Must identify dependencies
+   → Requires appropriate model capabilities
+
+2. Task Complexity: 8/10 (Granular work breakdown)
+   - Analyze story and break into tasks
+   - Determine appropriate task granularity
+   - Break tasks into atomic subtasks
+   - Identify task dependencies
+
+3. Context Understanding:
+   - Understand full story requirements
+   - Consider epic and project context
+   - Identify implicit implementation needs
+   → Requires appropriate model capabilities
+
+4. Consistency & Reliability:
+   - Task breakdown must be complete
+   - Dependencies must be accurate
+
+5. Speed Requirements: LOW (Secondary)
+   - Background processing
+   - Quality far more important
+
+6. Pricing Considerations: TERTIARY
+   - Single call per ceremony
+   - Critical for implementation planning
+
+RECOMMENDATION:
+
+Based on evaluation criteria above, select the model that best meets:
+1. PRIMARY: Output Quality - CRITICAL - Defines actual implementation plan
+2. SECONDARY: Task Complexity - Very High (8/10) - granular work breakdown with dependencies
+3. TERTIARY: Breakdown Capability - Exceptional atomic subtask identification
+4. Speed Requirements: LOW - Background processing, quality >>> speed
+5. Pricing: TERTIARY - Single call, critical impact justifies any tier
+
+Selection Guidance:
+- Prioritize models with exceptional hierarchical decomposition and breakdown capability
+- Require accurate identification of atomic, implementable subtasks
+- Must properly identify task dependencies
+- Must understand full story requirements in epic/project context
+- Completeness is critical (missing tasks cause implementation delays)
+- Reliability paramount (inconsistent granularity causes confusion)
+- Single call makes pricing tier less significant than quality
+
+Consider all available models from any provider (Claude, Gemini, OpenAI, etc.)`,
+    metadata: {
+      callFrequency: 1,
+      impact: 'CRITICAL',
+      taskComplexity: 8,
+      currentDefault: 'Claude Opus 4.6'
+    }
+  },
+
+  {
+    id: 'seed-validation',
+    ceremony: 'seed',
+    stage: 'validation',
+    stageName: 'Task Validation',
+    prompt: `TASK: Select optimal LLM model for task validation in seed ceremony
+PRIORITY: Best possible output quality
+
+CONTEXT:
+- Stage: validation (Task Validation)
+- Ceremony: seed
+- Purpose: AI validates task hierarchy completeness and feasibility
+- Input: Complete task/subtask hierarchy (3,000-10,000 tokens)
+- Output: Validation report identifying gaps and issues
+- Call frequency: ~20 calls per ceremony
+- Impact: VERY HIGH - Catches planning issues before development
+
+EVALUATION CRITERIA (BEST OUTPUT FIRST):
+
+1. Output Quality Requirements: VERY HIGH
+   - Must identify gaps in task coverage
+   - Must assess task granularity
+   - Must validate dependency correctness
+   → Requires appropriate model capabilities
+
+2. Task Complexity: 7/10 (Completeness validation + reasoning)
+   - Analyze task hierarchy for gaps
+   - Assess granularity appropriateness
+   - Validate dependency relationships
+   - Reason about implementation feasibility
+
+3. Context Understanding:
+   - Must understand full task hierarchy
+   - Cross-reference with story requirements
+   - Reason about dependencies
+   → Requires appropriate model capabilities
+
+4. Consistency & Reliability:
+   - Must consistently identify gaps
+   - Cannot miss critical completeness problems
+
+5. Speed Requirements: MODERATE (Secondary)
+   - 20 calls, parallel validation
+   - Quality far more important
+
+6. Pricing Considerations: MODERATE
+   - 20 calls = moderate volume
+   - Critical impact (prevents implementation issues)
+
+RECOMMENDATION:
+
+Based on evaluation criteria above, select the model that best meets:
+1. PRIMARY: Output Quality - VERY HIGH - Catches planning issues before development
+2. SECONDARY: Validation Reasoning - Sophisticated gap identification and granularity assessment
+3. TERTIARY: Task Complexity - High (7/10) - completeness validation with dependency reasoning
+4. Speed Requirements: MODERATE - 20 calls in parallel, quality >>> speed
+5. Pricing: MODERATE - 20 calls, critical impact justifies investment
+
+Selection Guidance:
+- Prioritize models with excellent completeness validation and gap identification
+- Require sophisticated reasoning about task granularity appropriateness
+- Must validate dependency correctness and implementation feasibility
+- Must cross-reference task hierarchy with story requirements
+- Cannot miss critical completeness problems (false negatives are costly)
+- Do not compromise on quality (prevents downstream implementation issues)
+
+Consider all available models from any provider (Claude, Gemini, OpenAI, etc.)`,
+    metadata: {
+      callFrequency: 20,
+      impact: 'VERY HIGH',
+      taskComplexity: 7,
+      currentDefault: 'Claude Sonnet 4.5'
+    }
+  },
+
+  {
+    id: 'seed-context-generation',
+    ceremony: 'seed',
+    stage: 'context-generation',
+    stageName: 'Task Context Generation',
+    prompt: `TASK: Select optimal LLM model for task context generation in seed ceremony
+PRIORITY: Best possible output quality
+
+CONTEXT:
+- Stage: context-generation (Task Context Generation)
+- Ceremony: seed
+- Purpose: AI generates context.md for each task
+- Input: Task with subtasks, story/epic/project contexts (1,500-4,000 tokens)
+- Output: Concise context.md with implementation approach (300-1,000 tokens)
+- Call frequency: ~10 calls per ceremony
+- Impact: VERY HIGH - Developers read immediately before implementing
+
+EVALUATION CRITERIA (BEST OUTPUT FIRST):
+
+1. Output Quality Requirements: VERY HIGH
+   - Must provide specific, actionable implementation guidance
+   - Technical details must be accurate
+   - Must highlight gotchas and edge cases
+   → Requires appropriate model capabilities
+
+2. Task Complexity: 7/10 (Implementation-focused technical context)
+   - Synthesize task requirements into guidance
+   - Provide appropriate technical detail
+   - Identify implementation approaches
+   - Balance brevity with actionability
+
+3. Context Understanding:
+   - Understand task within story/epic context
+   - Recognize relevant patterns
+   - Identify task dependencies
+   → Requires appropriate model capabilities
+
+4. Consistency & Reliability:
+   - Context must align with task requirements
+   - Technical guidance must be accurate
+
+5. Speed Requirements: MODERATE (Secondary)
+   - 10 calls, parallel generation
+   - Quality more important
+
+6. Pricing Considerations: MODERATE
+   - 10 calls = low-moderate volume
+   - Developer-facing, high impact
+
+RECOMMENDATION:
+
+Based on evaluation criteria above, select the model that best meets:
+1. PRIMARY: Output Quality - VERY HIGH - Developers read immediately before implementing
+2. SECONDARY: Technical Writing - Implementation-focused guidance with accuracy
+3. TERTIARY: Task Complexity - High (7/10) - synthesis with appropriate detail level
+4. Speed Requirements: MODERATE - 10 calls in parallel, quality >>> speed
+5. Pricing: MODERATE - 10 calls, developer-facing, high impact
+
+Selection Guidance:
+- Prioritize models with excellent technical writing for implementation guidance
+- Require ability to highlight gotchas, edge cases, and implementation approaches
+- Must provide specific, actionable guidance (not generic advice)
+- Must balance brevity with actionability (300-1,000 token output)
+- Technical details must be accurate (errors cause developer confusion)
+- Do not compromise on quality (directly impacts implementation efficiency)
+
+Consider all available models from any provider (Claude, Gemini, OpenAI, etc.)`,
+    metadata: {
+      callFrequency: 10,
+      impact: 'VERY HIGH',
+      taskComplexity: 7,
+      currentDefault: 'Claude Sonnet 4.5'
+    }
+  },
+
+  {
+    id: 'context-retrospective-documentation-update',
+    ceremony: 'context-retrospective',
+    stage: 'documentation-update',
+    stageName: 'Documentation Enhancement',
+    prompt: `TASK: Select optimal LLM model for documentation enhancement in context-retrospective ceremony
+PRIORITY: Best possible output quality
+
+CONTEXT:
+- Stage: documentation-update (Documentation Enhancement)
+- Ceremony: context-retrospective
+- Purpose: AI refines PROJECT.md based on implementation learnings
+- Input: PROJECT.md, git history, completed work (10,000-30,000 tokens)
+- Output: Updated PROJECT.md with refined descriptions and learnings
+- Call frequency: ~10 calls per ceremony
+- Impact: HIGH - Maintains PROJECT.md as source of truth
+
+EVALUATION CRITERIA (BEST OUTPUT FIRST):
+
+1. Output Quality Requirements: HIGH
+   - Must synthesize implementation learnings
+   - Technical updates must be accurate
+   - Must maintain documentation consistency
+   → Requires appropriate model capabilities
+
+2. Task Complexity: 7/10 (Technical synthesis and writing)
+   - Analyze git history
+   - Identify patterns and insights
+   - Synthesize into documentation updates
+
+3. Context Understanding:
+   - Must handle large context (PROJECT.md + git history)
+   - Understand implementation changes
+   - Identify obsolete information
+   → Large context window needed appropriate model capabilities
+
+4. Consistency & Reliability:
+   - Updates must not contradict existing content
+   - Technical accuracy critical
+
+5. Speed Requirements: LOW (Secondary)
+   - Background processing
+   - Quality more important
+
+6. Pricing Considerations: MODERATE
+   - 10 calls = low-moderate volume
+   - Important but not critical path
+
+RECOMMENDATION:
+
+Based on evaluation criteria above, select the model that best meets:
+1. PRIMARY: Output Quality - HIGH - Maintains PROJECT.md as source of truth
+2. SECONDARY: Technical Writing - Synthesis of implementation learnings
+3. TERTIARY: Task Complexity - High (7/10) - git history analysis + documentation
+4. Speed Requirements: LOW - Background processing, quality >>> speed
+5. Pricing: MODERATE - 10 calls, important but not critical path
+
+Selection Guidance:
+- Prioritize models with excellent technical writing and documentation synthesis
+- Require ability to analyze git history and identify implementation patterns
+- Must maintain consistency with existing documentation (no contradictions)
+- Must handle large context (PROJECT.md + git history = 10K-30K tokens)
+- Technical accuracy critical (documentation is project source of truth)
+- Consider balance of quality vs pricing for moderate-volume task
+
+Consider all available models from any provider (Claude, Gemini, OpenAI, etc.)`,
+    metadata: {
+      callFrequency: 10,
+      impact: 'HIGH',
+      taskComplexity: 7,
+      currentDefault: 'Claude Sonnet 4.5'
+    }
+  },
+
+  {
+    id: 'context-retrospective-context-refinement',
+    ceremony: 'context-retrospective',
+    stage: 'context-refinement',
+    stageName: 'Context Enhancement',
+    prompt: `TASK: Select optimal LLM model for context refinement in context-retrospective ceremony
+PRIORITY: Best possible output quality
+
+CONTEXT:
+- Stage: context-refinement (Context Enhancement)
+- Ceremony: context-retrospective
+- Purpose: AI enhances context.md with implementation insights
+- Input: context.md, git diffs, PRs, issue comments (5,000-15,000 tokens)
+- Output: Enhanced context.md with insights, patterns, code examples
+- Call frequency: ~15 calls per ceremony
+- Impact: VERY HIGH - Enhanced context improves all future AI agent performance
+
+EVALUATION CRITERIA (BEST OUTPUT FIRST):
+
+1. Output Quality Requirements: VERY HIGH
+   - Must extract valuable insights from code changes
+   - Must identify implementation patterns
+   - Should include practical code examples
+   → Requires appropriate model capabilities
+
+2. Task Complexity: 8/10 (Code analysis + insight extraction)
+   - Analyze code diffs
+   - Identify patterns and best practices
+   - Extract insights from issue discussions
+   - Synthesize into actionable context
+
+3. Context Understanding:
+   - Understand code changes in project context
+   - Analyze git diffs
+   - Synthesize from multiple sources
+   → Requires appropriate model capabilities
+
+4. Consistency & Reliability:
+   - Enhanced context must align with original
+   - Code examples must be accurate
+   → Both appropriate model capabilities
+
+5. Speed Requirements: LOW (Secondary)
+   - Background processing
+   - Quality far more important
+
+6. Pricing Considerations: MODERATE
+   - 15 calls = moderate volume
+   - Very high leverage (improves all future AI work)
+
+RECOMMENDATION:
+
+Based on evaluation criteria above, select the model that best meets:
+1. PRIMARY: Output Quality - VERY HIGH - Enhances context for all future AI agents
+2. SECONDARY: Code Analysis - Exceptional insight extraction from git diffs
+3. TERTIARY: Task Complexity - Very High (8/10) - pattern identification + synthesis
+4. Speed Requirements: LOW - Background processing, quality >>> speed
+5. Pricing: MODERATE - 15 calls, very high leverage (improves all future ceremonies)
+
+Selection Guidance:
+- Prioritize models with exceptional code analysis and pattern recognition
+- Require ability to extract valuable insights from code diffs and PRs
+- Must identify implementation patterns, gotchas, and best practices
+- Must synthesize from multiple sources (git diffs, issues, comments)
+- Code examples must be accurate (errors propagate to future AI work)
+- Very high leverage justifies premium tier (enhanced context used in all future work)
+- Quality directly impacts all future AI agent effectiveness
+
+Consider all available models from any provider (Claude, Gemini, OpenAI, etc.)`,
+    metadata: {
+      callFrequency: 15,
+      impact: 'VERY HIGH',
+      taskComplexity: 8,
+      currentDefault: 'Claude Opus 4.6'
+    }
+  }
+];
+
+/**
+ * Get evaluation prompt by ID
+ * @param {string} id - Prompt ID (e.g., 'sponsor-call-suggestions')
+ * @returns {Object|null} Evaluation prompt object or null if not found
+ */
+export function getPromptById(id) {
+  return EVALUATION_PROMPTS.find(p => p.id === id) || null;
+}
+
+/**
+ * Get all prompts for a specific ceremony
+ * @param {string} ceremony - Ceremony name (e.g., 'sponsor-call')
+ * @returns {Array} Array of evaluation prompts for the ceremony
+ */
+export function getPromptsByCeremony(ceremony) {
+  return EVALUATION_PROMPTS.filter(p => p.ceremony === ceremony);
+}
+
+/**
+ * Get prompt statistics
+ * @returns {Object} Statistics about evaluation prompts
+ */
+export function getPromptStats() {
+  const ceremonies = [...new Set(EVALUATION_PROMPTS.map(p => p.ceremony))];
+  const totalCalls = EVALUATION_PROMPTS.reduce((sum, p) => sum + p.metadata.callFrequency, 0);
+
+  return {
+    totalPrompts: EVALUATION_PROMPTS.length,
+    ceremonies: ceremonies.length,
+    ceremonyList: ceremonies,
+    estimatedTotalCalls: totalCalls,
+    impactDistribution: {
+      CRITICAL: EVALUATION_PROMPTS.filter(p => p.metadata.impact === 'CRITICAL').length,
+      'VERY HIGH': EVALUATION_PROMPTS.filter(p => p.metadata.impact === 'VERY HIGH').length,
+      HIGH: EVALUATION_PROMPTS.filter(p => p.metadata.impact === 'HIGH').length,
+      MEDIUM: EVALUATION_PROMPTS.filter(p => p.metadata.impact === 'MEDIUM').length
+    }
+  };
+}