skills.rest
Jin9
Jin9COMMUNITY
Explore all skills by this author →
GitHub

model-selection

Assigns models and reasoning effort per agent role using privacy gates and cost-per-task criteria.

Updated Apr 26, 2026
Search Tags:#multi-agent#cost optimization#llm routing#agentic workflow#model selection#fallback#reasoning effort
One-click install
npx skills add https://github.com/Jin9/skillify-foundation --skill model-selection-jin9
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: model-selection
Source: https://github.com/Jin9/skillify-foundation/tree/main/treasury/model-selection
Command: npx skills add https://github.com/Jin9/skillify-foundation --skill model-selection-jin9

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Teams building multi-agent workflows often overspend by running one frontier model everywhere or underperform by picking models on token price alone. This Skill produces a criteria-driven model-assignment policy that matches each agent role to the right capability tier, reasoning effort, and fallback strategy. ## Core Features & Use Cases - Privacy-First Gating: Classifies data sensitivity per role and eliminates ineligible models (hosted-standard, ZDR/BAA, VPC, self-hosted open-weight) before any capability comparison. - Per-Role Assignment Matrix: Maps each agent to a capability tier, reasoning effort, context budget, structured decoding support, cost posture, and fallback with rationale. - Cost-Per-Successful-Task Budgeting: Optimizes CPST (including retries, tool loops, and caching) instead of raw token price, with routing regime and fallback topology specified. - Use Case: For a 13-stage payments delivery squad, the skill gates regulated data to frontier-closed models, assigns frontier-reasoning to orchestrator/reviewer roles, mid-tier to coding, and cheap-fast with strict JSON decoding to extraction — with a quarterly re-evaluation date. ## Quick Start Ask the agent to assign a model and reasoning effort to each agent role in your workflow, given each role's data sensitivity and latency budget.

Frequently Asked Questions about model-selection

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I choose which LLM each agent in my workflow should use?▼

Classify the data sensitivity each role handles first, then assign the cheapest model that clears that role's capability bar. Use frontier-reasoning tiers for planning and review, mid-tier for implementation, and cheap-fast tiers for extraction and summarization.

Should I use one frontier model for every agent role?▼

No. Running one model in every role overspends on simple tasks and underperforms on hard ones. Match capability to the role and set reasoning effort per task difficulty, since mid-tier models with high reasoning effort often beat frontier models with zero effort on logic tasks.

How do I budget LLM costs for an agentic workflow?▼

Budget in cost-per-successful-task (CPST), not token price. Account for retries, tool loops, and super-linear context growth, and exhaust prompt caching with strict prefix discipline and batching before considering any model swap.

Does this skill route or proxy LLM requests at runtime?▼

No. It produces a model-assignment policy for human sign-off and does not operate a routing gateway or LLM proxy. Runtime routing belongs to tools like OpenRouter, RouteLLM, or LiteLLM.

How should I handle model fallback in a multi-agent system?▼

Specify a fallback per role with an alternate model, defaulting to same-provider cross-region topology. Keep retries idempotent, honor Retry-After headers, distinguish terminal from retryable errors, and always log the actually-served model.

When should I not use model tiering as a safety control?▼

Never treat a stronger model as a safety control for high-stakes or autonomous actions. Prompt injection is unmitigated at the model layer, so such actions require architectural guardrails independent of model choice.

Related Skills

Explore More Software Engineering Skills→
itallstartedwithaideaitallstartedwithaidea
31

multi-model-routing

Route tasks to model providers by cost, latency, and circuit-breaker fallback.

Community
Advanced
weirdgmeweirdgme
4

observability-engineering

Instrument distributed systems with OpenTelemetry and correlate logs, metrics, and traces.

Community
Advanced
Nairon-AINairon-AI
7

flux-threat-model

Detect and document security threats in code repositories using STRIDE methodology.

Official
Advanced
Randi-AgentRandi-Agent
1

openrouter-llm

Route model requests across multiple providers through the OpenRouter API gateway.

Official
Advanced