Разработка и оптимизация промптов для LLM

ai-tooling

Prompt Engineer is a Claude Code skill that writes, refactors, and evaluates prompts for LLMs — producing optimized templates, structured output schemas, evaluation rubrics, and test suites. Reach for it when model outputs are inconsistent and it's unclear why; when an existing prompt needs to be ported to a different model without quality loss; when the use case demands strict JSON output or function calling; or when you want to add chain-of-thought reasoning or few-shot examples but aren't sure how to structure them correctly. The skill covers zero-shot, few-shot, CoT, ReAct, and tree-of-thoughts patterns; system prompt design with personas, guardrails, and injection defense; context management with attention budgets and degradation mitigation; and evaluation frameworks with accuracy and consistency metrics. It follows a disciplined loop — requirements analysis, prompt design, diverse test-case evaluation, single-change iteration, and versioning — with a built-in checkpoint that surfaces failure patterns before proceeding if test-set accuracy falls below 80%. Model-specific guidance is included for GPT-4, Claude, and Gemini families. The scope is prompt design and optimization, not task orchestration (use rag-architect for retrieval pipeline work).