Skill

ai-llm-patterns

AI/LLM integration patterns: Anthropic SDK streaming with TypeScript, RAG architecture (chunking, embedding, vector search with pgvector), tool use / function calling, model selection guide (Haiku 4.5 vs Sonnet 4.6 vs Opus 4.6), prompt caching, structured output with Zod, token cost management. Use when building AI features, RAG pipelines, AI agents.

Install

npx claudepluginhub rnavarych/alpha-engineer --plugin billy-milligan

Tool Access

This skill is limited to using the following tools:

ReadGrepGlob

Preview

- Building AI features with Claude / Anthropic SDK

Supporting Assets

references/llm-ops.mdreferences/prompt-engineering.mdreferences/rag-architecture.md

SKILL.md

Similar Skills

agent-eval

Compares coding agents like Claude Code and Aider on custom YAML-defined codebase tasks using git worktrees, measuring pass rate, cost, time, and consistency.

everything-claude-code

156.2k

agent-harness-construction

Designs and optimizes AI agent action spaces, tool definitions, observation formats, error recovery, and context for higher task completion rates.

everything-claude-code

156.2k

accessibility

Designs, implements, and audits WCAG 2.2 AA accessible UIs for Web (ARIA/HTML5), iOS (SwiftUI traits), and Android (Compose semantics). Audits code for compliance gaps.

everything-claude-code

156.2k

Stats

Parent Repo Stars9

Parent Repo Forks0

Last CommitMar 5, 2026

Actions

View Source View Plugin View on GitHub View README

ai-llm-patterns

Core Principles

Model selection matters for cost — Haiku is 10× cheaper than Sonnet; Sonnet is 5× cheaper than Opus; use the smallest model that can do the task

Streaming for user experience — start showing text immediately; never make users wait for full response

Tool use for deterministic operations — LLM decides what to do; your code does it safely; never trust LLM output directly for DB writes

Prompt caching for repeated context — system prompts, documents, RAG context: cache if >1024 tokens

RAG chunk size matters — too small (50 tokens) = no context; too large (2000 tokens) = irrelevant context dilutes relevance

References available

references/prompt-engineering.md — streaming with SSE, model selection guide (Haiku/Sonnet/Opus), prompt caching, structured output with Zod

references/rag-architecture.md — chunking strategy, embedding with text-embedding-3-small, pgvector cosine search, retrieval pipeline

references/llm-ops.md — tool use / agent loop pattern, cost tracking, token budgeting, multi-model routing, error handling

AI/LLM Integration Patterns

When to Use This Skill

Building AI features with Claude / Anthropic SDK
Implementing RAG (Retrieval-Augmented Generation)
Using tool use / function calling patterns
Choosing the right Claude model for the task
Managing token costs and prompt caching

Core Principles

Model selection matters for cost — Haiku is 10× cheaper than Sonnet; Sonnet is 5× cheaper than Opus; use the smallest model that can do the task
Streaming for user experience — start showing text immediately; never make users wait for full response
Tool use for deterministic operations — LLM decides what to do; your code does it safely; never trust LLM output directly for DB writes
Prompt caching for repeated context — system prompts, documents, RAG context: cache if >1024 tokens
RAG chunk size matters — too small (50 tokens) = no context; too large (2000 tokens) = irrelevant context dilutes relevance

References available

references/prompt-engineering.md — streaming with SSE, model selection guide (Haiku/Sonnet/Opus), prompt caching, structured output with Zod
references/rag-architecture.md — chunking strategy, embedding with text-embedding-3-small, pgvector cosine search, retrieval pipeline
references/llm-ops.md — tool use / agent loop pattern, cost tracking, token budgeting, multi-model routing, error handling