skills.rest
nebius
nebiusOFFICIAL
Explore all skills by this author →
WebsiteGitHub

ai-stack

Selects and reviews AI architecture components across models, runtimes, retrieval, and operations.

7|5|Updated Jul 22, 2025
Search Tags:#decision record#llm#agent framework#mcp#inference#technology selection#ai architecture
One-click install
npx skills add https://github.com/nebius/nebius-ps-services --skill ai-stack-nebius
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: ai-stack
Source: https://github.com/nebius/nebius-ps-services/tree/main/skills/ai-stack
Command: npx skills add https://github.com/nebius/nebius-ps-services --skill ai-stack-nebius

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve? Choosing AI components for a system often leads to over-engineered stacks with unjustified frameworks, gateways, and databases. This Skill applies hard gates, evidence states, and workload contracts to select or review only the AI layers a concrete workload actually requires. ## Core Features & Use Cases - Workload-driven selection: Freezes per-workload contracts with acceptance thresholds, then classifies each capability as deterministic code, direct model call, deterministic workflow, or agent before naming any product. - Gated component decisions: Assigns every component a status (Required, Conditional, Deferred, Rejected) and every claim an evidence state (Measured, Officially documented, Assumed), covering training, inference, agents, MCP, retrieval, evaluation, and safety. - Use Case: When designing a new LLM-powered service, use this Skill to decide whether you need an agent framework like Pydantic AI or LangGraph, a durable workflow engine, a vector database, or simply a direct provider API call, with rejected alternatives and switch conditions recorded. ## Quick Start Use the ai-stack skill to select the AI components for a customer-support chatbot workload with latency and cost acceptance gates.

Frequently Asked Questions about ai-stack

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

FAQPage Schema
How do I choose an AI agent framework for a Python application?▼

Classify the capability first: use a direct model call or deterministic workflow when application code knows the steps, and an agent only when the model must choose actions or continuation. For genuinely agentic portable Python workloads, Pydantic AI Agent is the default; LangGraph applies only when explicit graph semantics are required.

When should I use MCP versus A2A for agent interoperability?▼

Use MCP for independently deployed reusable tools, resources, or capabilities consumed by agent hosts. Use A2A only when the remote unit is an independently owned and deployed agent; a tool or deterministic service should remain an ordinary API or MCP capability.

Do I need a vector database for my RAG application?▼

Not by default. The retrieval ladder starts with deterministic source access, then PostgreSQL full-text search, then pgvector after a labeled relevance gate passes. A dedicated vector system like Qdrant is justified only by scale, filter behavior, or ownership requirements.

What is the difference between ai-stack and ai-agent-design?▼

ai-agent-design owns agent-subsystem behavior classification, topology, authority, contracts, memory, and governance. ai-stack consumes that frozen contract and selects the concrete technologies and components; it never reopens the behavior decisions.

When should I add a durable workflow engine like Temporal?▼

Add one qualified durable workflow owner only when the workload requires timers, long approvals, external side effects, compensation, or cross-service recovery. Compare Temporal, DBOS, Prefect, Restate, or an existing engine against the workload rather than mandating one.

Why does the skill reject vendor benchmark claims?▼

Vendor documentation is not target measurement. Every material claim must be Measured on a representative setup, Officially documented by a primary source, or marked Assumed, and an Assumed claim cannot close a hard gate.

Related Skills

Explore More Software Engineering Skills→
alvisalvis
4

write-skill

Create, update, and verify Agent Skills for Claude Code, Codex, and Grok Build.

Community
Advanced
korwlik23korwlik23

brainstorming

Turns rough ideas into approved implementation plans before code is written.

Community
Advanced
CellCogCellCog
6

research-cog

Conduct citation-backed research across competitive, market, and academic topics.

Official
Advanced
akvoakvo
2

bmad-architect

Design system architectures and document decisions with Architecture Decision Records.

Official
Advanced