Optimize LLM Prompts for Production

Prompt Optimization is a skill for transforming basic instructions into production-ready prompts using techniques such as chain-of-thought reasoning, few-shot learning, constitutional AI self-critique, and model-specific formatting. Per the skill's own framing, effective prompt engineering can improve accuracy by 40%, reduce hallucinations by 30%, and cut costs by 50-80% through token optimization. It gives an eight-step process: analyzing the current prompt against a clarity/structure/model-alignment/performance framework; applying chain-of-thought enhancement through standard CoT, zero-shot CoT, and tree-of-thoughts patterns; implementing few-shot learning with strategic example selection that includes an explicit error case; applying constitutional AI self-critique loops that check accuracy, safety, and quality; model-specific optimization templates for GPT-5/GPT-4o using hash-delimited sections, Claude 4.5/4 using XML-tagged context/task/thinking/output_format blocks, and Gemini Pro/Ultra using bolded headers; RAG-integrated prompts that instruct citation of retrieved sources and disclosure of coverage gaps; an evaluation framework with a 20-case test protocol split across typical, edge, adversarial, and out-of-scope cases, plus an LLM-as-judge rubric scoring task completion, accuracy, reasoning, format, and safety; and production deployment guidance covering prompt versioning with a canary/staged rollout and a rollback threshold, plus structured error-handling templates for insufficient information, contradictions, limitations, and safety concerns. It includes four worked before/after examples - customer support, data analysis, code generation, and a meta-prompt generator - and specifies an output format of an optimized prompt plus a YAML optimization report covering original assessment, improvements applied, performance projection, testing recommendations, and deployment strategy.

Use it when working on prompt optimization tasks or workflows, or needing guidance, best practices, or checklists for prompt optimization. Do not use it when the task is unrelated to prompt optimization or needs a different domain or tool outside this scope.

Input is a basic or underperforming prompt plus its target model and use case. Output is a comprehensive optimization report: the complete production-ready prompt with all enhancements applied, a YAML analysis of strengths, weaknesses, and token count, a list of improvements applied with their measured impact, a performance projection covering success rate, token efficiency, quality, and safety, testing recommendations calling for LLM-as-judge evaluation with human validation across 20 test cases and a 48-hour A/B test, and a deployment strategy naming model, temperature, and max tokens. It closes with five usage guidelines: use the optimized prompt exactly as delivered, apply the recommended parameters, run the test cases before production, monitor metrics after launch, and iterate based on performance data.

Model-specific templates target GPT-5/GPT-4o, Claude 4.5/4, and Gemini Pro/Ultra by name, and the RAG pattern integrates with a document-retrieval step that feeds retrieved context into the prompt for citation-backed synthesis. The example deployment strategy recommends a temperature of 0.7 and a 2000-token cap when serving the optimized prompt.

Prompt engineers, AI application developers, and teams building or hardening LLM-backed features who need a repeatable process for turning a rough instruction into a tested, production-ready prompt.