prompt-craftlisted

Writing, improving, and auditing prompts for LLMs (Claude 4.x/5, GPT-5.x, Gemini 3, Perplexity Sonar and similar) according to modern prompt- and context-engineering practice of 2025–2026. Use this skill whenever the user asks to "write a prompt", "draft a prompt", "make a system prompt", "a prompt for an agent", "a prompt for GPT/Claude/Gemini/Perplexity", "improve my prompt", "rewrite this prompt", "make it jailbreak-resistant", "protect the prompt from injection", "write instructions for a model", "prompt template", "few-shot examples", "system prompt", "configure reasoning/thinking", "why does the model ignore my instructions", or when the user hands over a raw task meant for another model and wants a polished instruction back. Trigger even if the word "prompt" is never said — if the user is preparing an instruction to feed to another LLM system (a chatbot, an agent, a RAG pipeline, a classifier, an LLM judge), that is a prompt-craft task.
kravcovtech/prompt-craft · ★ 0 · AI & Automation · score 60

Install: claude install-skill kravcovtech/prompt-craft

# Prompt Craft A skill for producing high-quality prompts for modern LLMs (Claude 4.x/5, GPT-5.x, Gemini 3, Perplexity Sonar and similar). It draws on the official 2025–2026 practices of Anthropic, OpenAI, Google, and Perplexity, and on research (Wharton Prompting Science, Lost in the Middle, Context Rot, GEO). The core of the approach: first determine the model class (reasoning vs. standard), then the platform, then assemble the prompt from the canonical structure and run it through a checklist. ## When to use this skill - The user asks for a prompt written from scratch - The user brings a draft prompt and asks to improve/rewrite/simplify it - The user complains that a model is doing something wrong and the prompt needs fixing - The user describes a task for an agent, classifier, RAG system, generator, or judge and wants an instruction for the model - The user asks for few-shot examples, a system prompt, an instruction template, or an output JSON schema - The user asks about protecting a prompt from injection/jailbreaks, or about system-prompt leakage If the task is simply to answer the user as Claude (rather than to prepare an instruction for another LLM session), the skill isn't needed. --- ## Basic workflow 1. **Clarify the intent** (if it isn't obvious). Ask no more than 1–3 questions, and only if you genuinely can't proceed without them. If you can reasonably guess — guess, do the work, and note your assumptions at the end. 2. **Determine the model class: reasoni