analysis-orchestration

Model selection, cost estimation, and batch processing setup for AI-assisted coding. Helps researchers configure their analysis approach with awareness of tradeoffs.

When to Use

Use this skill when:

• User asks about which AI model to use for coding
• User mentions "cost", "batch", "API", or "configure analysis"
• User wants to process multiple documents
• User needs to understand model capabilities and costs
• Starting Stage 2 and needing to plan the approach

Capabilities

1. Model Selection Guidance - Help choose between models based on task needs
2. Cost Estimation - Estimate API costs before processing
3. Batch Strategy - Plan efficient document processing
4. API Configuration - Set up for programmatic coding (future)

Model Selection Guide

For Document Coding (Stage 2)

Model	Best For	Cost	Quality
Claude Opus 4.5	Complex interpretive coding, nuanced themes	$$$	Highest
Claude Sonnet 4	Balanced quality and cost for systematic coding	$$	High
Claude Haiku	Initial passes, high volume, simple categorization	$	Good

Recommendation by Task

Deep Interpretive Coding (@dialogical-coder)

• Use: Opus 4.5 or Sonnet 4
• Why: Requires nuanced understanding, theoretical sensitivity
• Pattern: Process 5-10 documents per session with reflection

Initial Categorization

• Use: Sonnet 4 or Haiku
• Why: Applying established codes is less interpretively demanding
• Pattern: Batch process with human review

Pattern Characterization

• Use: Opus 4.5
• Why: Requires integration across documents, theoretical abstraction

Cost Estimation

Rough Estimates (2025 pricing)

Documents	Model	Estimated Cost
10 interviews (~50 pages)	Opus	$15-25
10 interviews (~50 pages)	Sonnet	$5-10
50 documents	Opus	$75-125
50 documents	Sonnet	$25-50

Variables:

• Document length (tokens)
• Coding depth (passes per document)
• Output verbosity (full reasoning vs brief)

Scripts

estimate-costs.js

Estimates API costs based on document characteristics.

node skills/analysis-orchestration/scripts/estimate-costs.js \
  --documents 25 \
  --avg-pages 5 \
  --model sonnet \
  --passes 2

Returns: Estimated cost range and token counts.

Batch Processing Strategy

Small Corpus (10-30 documents)

• Process individually with full dialogical coding
• High engagement, rich reasoning
• Best for: Interpretive research, theory building

Medium Corpus (30-100 documents)

• Batch in groups of 10
• First pass: categorization
• Second pass: deep coding on interesting cases
• Best for: Mixed-methods, systematic reviews

Large Corpus (100+ documents)

• Strategic sampling for deep coding
• Batch categorization with random quality checks
• Best for: Large-scale studies, triangulation

Decision Trees

Which Model Should I Use?

Is this initial exploratory coding?
├── Yes → Consider Haiku for volume, validate with Sonnet
└── No, this is interpretive coding
    ├── Budget constrained?
    │   ├── Yes → Sonnet 4 (good balance)
    │   └── No → Opus 4.5 (best quality)
    └── Need to process >50 documents?
        ├── Yes → Two-pass: Haiku then Sonnet on subset
        └── No → Single-pass with Sonnet or Opus

How Many Documents Per Session?

Using @dialogical-coder (4-stage process)?
├── Yes → 5-10 documents per session (reflection breaks)
└── No, systematic application?
    ├── Complex coding scheme → 10-15 documents
    └── Simple categorization → 20-30 documents

Integration with Interpretive Orchestration

Stage 1

• No AI models needed (manual coding)
• Focus on human theoretical sensitivity

Stage 2 Phase 1 (Parallel Streams)

• Stream A (theoretical): Sonnet for literature analysis
• Stream B (empirical): Start with Sonnet, elevate complex cases to Opus

Stage 2 Phase 2 (Synthesis)

• Opus recommended for integration work
• Cross-stream synthesis requires nuanced reasoning

Stage 2 Phase 3 (Pattern Characterization)

• Opus for pattern identification
• Sonnet for validation passes

Related

• Commands: /qual-configure-analysis triggers this skill
• Agents: @research-configurator provides interactive guidance
• Skills: coding-workflow/ for batch processing execution