Engineer optimized prompts for LLM reasoning and output

This skill turns an agent into a prompt-engineering expert across GPT, Claude, Llama, and other transformer-based models. It covers core principles - clarity and specificity, explicit role/persona definition, and structured step-by-step reasoning - and layers in advanced techniques: chain-of-thought prompting that breaks a problem into identify-analyze-relate-synthesize steps, few-shot learning with input/output example pairs, and a tree-of-thoughts framework that explores multiple reasoning paths (with pros/cons for each) before selecting the best one.

It also covers prompt optimization: temperature guidance mapped to task type (0.1-0.3 for factual/code tasks, 0.4-0.7 for creative/balanced work, 0.8-1.0 for highly creative ideation), token-efficiency rewriting (turning a verbose paragraph request into a bulleted, constrained one), and error-prevention constraints such as instructing the model not to guess at missing information and to flag uncertainty explicitly. Domain-specific templates are included for code generation (functionality, style guide, error handling, usage examples), data analysis (summary statistics, patterns, anomalies, confidence levels), and creative writing (setting, characters, tone, length, theme).

Use this skill when designing, debugging, or optimizing prompts for an LLM-based feature: picking a reasoning technique (chain-of-thought vs. few-shot vs. tree-of-thoughts), tuning temperature for the task, converting a vague prompt into a structured one, or setting up an A/B test between two prompt versions. It also covers model-specific tuning - GPT responds well to direct instructions and explicit role assignment, Claude performs well with constitutional-AI-style and collaborative framing, and open models like Llama or Mistral often need more explicit instruction templates. It is a technique and template reference, not a prompt-management platform - it does not itself store, version, or deploy prompts; it documents the discipline (clear versioning, documented changes, cross-model testing) for doing that yourself.

Input: a description of the LLM task (classification, generation, analysis, creative writing) and which model family it targets. Output: a structured prompt template with role definition, reasoning framework, output-format constraints, and - where relevant - few-shot examples or a chain-of-thought/tree-of-thoughts scaffold, plus guidance on temperature setting and a checklist for debugging common failure modes (vague outputs, inconsistent formatting, off-topic responses, incomplete reasoning).

The testing pattern shown is a small Python structure (prompt_tests list of dicts with prompt, expected_elements, and success_criteria) for automated prompt-output verification; the rest of the skill is model-agnostic prompt text applicable to any LLM API.

Developers and prompt engineers building LLM-powered features who need concrete, reusable prompt patterns - chain-of-thought, few-shot, tree-of-thoughts, temperature tuning, and model-specific adjustments - rather than a general explanation of what prompt engineering is.