Standards & Patterns
Model & Routing Catalog
Maintained by: Ada (Agentic Pattern Designer) + Riley (R&D Analyst) Last updated: 2026-03-29 Decision: TFD-0014 (Component Taxonomy — Category 3)
Overview
This catalog documents which LLM handles which task across all digital talent production. Model selection balances cost, quality, privacy, and latency — the right model for the right job.
Every skill shipped with a digital talent includes a model recommendation. This catalog provides the selection framework that drives those recommendations.
Available Models
Claude Family (Primary — Anthropic)
| Model | Tier | Strengths | Cost | Context | When to Use |
|---|---|---|---|---|---|
| Claude Opus | Premium | Complex reasoning, orchestration, multi-step planning, nuanced judgment | $$$ | 200K | Orchestrators, planners, complex analysis, architecture decisions |
| Claude Sonnet | Standard | Balanced analysis, writing, code generation, standard workflows | $$ | 200K | Most skill execution, document generation, standard analysis |
| Claude Haiku | Economy | Fast execution, template filling, validation, simple lookups | $ | 200K | File operations, format validation, simple transformations, evaluators |
Model Capability Boundaries
| Capability | Haiku | Sonnet | Opus |
|---|---|---|---|
| Template filling | Excellent | Excellent | Overkill |
| Document generation | Basic | Excellent | Excellent |
| Multi-step reasoning | Limited | Good | Excellent |
| Architecture analysis | Not recommended | Good | Excellent |
| Orchestration/planning | Not recommended | Acceptable | Excellent |
| Code generation | Basic | Excellent | Excellent |
| Self-evaluation | Unreliable | Acceptable | Good (but see Principle 1) |
| Domain expertise | Limited | Good | Excellent |
Local Models (Privacy Routing — Future)
| Model | Use Case | Status |
|---|---|---|
| Local LLM (e.g., Llama, Mistral) | Sensitive data that cannot leave client infrastructure | R&D — not yet evaluated |
| On-prem deployment | Regulated industries (healthcare, finance, government) | R&D — requires AI Model Expert role |
Gap: No formal evaluation of local models. Deferred pending AI Model Expert role creation (see deferred decisions).
Selection Criteria
Primary Decision: Task Complexity
Is this task simple execution (template fill, file ops, validation)?
YES → Haiku
NO → Does it require multi-step reasoning or orchestration?
YES → Opus
NO → Sonnet (default for most skills)
Secondary Factors
| Factor | Guidance | Override Direction |
|---|---|---|
| Privacy | Sensitive client data → local model (when available); otherwise document risk | May force model change regardless of task complexity |
| Cost sensitivity | Client on budget tier → Sonnet ceiling, Haiku preferred | Downgrade from default |
| Latency | Interactive/streaming tasks → Haiku or Sonnet; batch OK with Opus | Faster model for UX |
| Quality threshold | Output directly faces client stakeholders → upgrade one tier | Upgrade from default |
| Evaluation tasks | Separate evaluator agent → can use Haiku with structured criteria | Downgrade: evaluators don't need to generate, just judge |
Cost Optimization Patterns
| Pattern | How It Works | When to Apply |
|---|---|---|
| Start high, downgrade | Begin with Sonnet for all skills, downgrade to Haiku after validation shows quality holds | Default approach during assembly |
| Opus for planning only | Reserve Opus for orchestrator/planner; generators use Sonnet, evaluators use Haiku | Multi-agent architectures (future) |
| Haiku for validation | Use Haiku for all verification/validation steps where criteria are explicit | Anywhere quality criteria can be formalized as a checklist |
| Batch processing | Use Opus for complex batch tasks where latency doesn't matter but quality does | Architecture reviews, compliance audits |
Per-Skill Assignment Guide
During assembly (Phase C), each skill gets a model recommendation. Use this mapping:
Typical Assignments by Skill Type
| Skill Type | Default Model | Upgrade To | Downgrade To |
|---|---|---|---|
| Orientation / analysis | Sonnet | Opus (if domain is complex) | — |
| Document generation | Sonnet | Opus (if high-stakes deliverable) | Haiku (if template-driven) |
| Diagram generation | Sonnet | — | Haiku (if template + data fill) |
| Validation / verification | Haiku | Sonnet (if judgment needed) | — |
| Orchestrator / workflow | Opus | — | Sonnet (if workflow is linear) |
| Data transformation | Haiku | Sonnet (if complex mapping) | — |
| Requirements capture | Sonnet | Opus (if client interviews) | Haiku (if form-based) |
| Search / lookup | Haiku | — | — |
| Quality evaluation | Haiku + criteria | Sonnet (if subjective judgment) | — |
| Client-facing report | Sonnet | Opus (if executive audience) | — |
Work Order Override
Clients can specify model preferences in the work order (Section 7: Model Selection). Client preferences override the defaults above when:
- Client pays for premium tier → Opus ceiling unlocked for all skills
- Client on budget → Sonnet ceiling, Haiku preferred
- Client has privacy requirements → local model routing (when available)
Model Routing Architecture
Current: Static Assignment
Today, model assignment is static — set once during assembly and baked into each skill's frontmatter. The CLAUDE.md skills table documents the assignment.
# In skill frontmatter
model: sonnet
# In CLAUDE.md skills table
| Skill | Command | Input | Output | Model |
|-------|---------|-------|--------|-------|
| Orientation | /orientation | request folder | analysis.md | sonnet |
Future: Dynamic Routing (R&D)
With Agent SDK, model routing could become dynamic — the orchestrator decides which model to use based on task complexity at runtime:
| Scenario | Routing Decision |
|---|---|
| Simple request, known pattern | Haiku generates, Haiku validates |
| Standard request | Sonnet generates, Haiku validates |
| Complex request, novel domain | Opus plans, Sonnet generates, Sonnet validates |
| High-stakes deliverable | Opus plans, Opus generates, Sonnet validates |
This requires Agent SDK evaluation (R&D intake logged 2026-03-29).
Update Triggers
| Trigger | Action |
|---|---|
| New Claude model release (e.g., Claude 5) | Riley evaluates; Ada updates capability boundaries and selection criteria |
| Pricing change | Ada updates cost tier and optimization patterns |
| Client requests local model | Escalate to R&D; flag AI Model Expert role need |
| Production feedback (model underperforms) | Pablo reports; Ada adjusts assignment for that skill type |
| New model family available | Riley evaluates through R&D pipeline (TFD-0012) |
References
- Assembly guide Phase C, Step 5:
production-lines/digital-talent/assembly.md - CLAUDE template Section 7:
production-lines/digital-talent/templates/CLAUDE-template.md - Skill template:
production-lines/digital-talent/templates/skill-template.md - Harness design principles:
docs/superpowers/specs/2026-03-29-component-taxonomy-design.md(Section 3.2) - Technology radar:
departments/executive/rd-analyst-riley/technology-radar.md