Prompt Alchemist

The Problem: From "Guessing" to "Systematic Optimization"

Engineers and content creators often struggle when optimizing prompts for large language models. We intuitively write instructions like "make it professional" or "analyze this for me," expecting high-quality outputs, but the results are often vague or poorly formatted. The root cause is that the prompt itself is ambiguous: it lacks clear structure, doesn't define a role or output format, and fails to guide the model into deep thinking. This leads to inefficient, iterative debugging and misunderstandings of model capabilities.

Core Capability: The Four-Element Distillation Framework & Master Methodologies

At its core is a systematic framework called "Four-Element Distillation," designed to refine your raw initial prompt into a precise, structured, and reusable version. The workflow involves four key steps:

1. 🔍 Diagnose: It employs Schulhoff's classification and Li Jigang's Johari Window model to "diagnose" the original prompt. It identifies the task type (e.g., code generation, reasoning), locates knowledge gaps (whether it's "Human Knows/AI Knows" or "Human Doesn't Know/AI Doesn't Know"), and pinpoints structural flaws like role stacking or missing formats.
2. 🏗️ Structure: Drawing from Riley Goodside's experimental approach, it prioritizes structure over wording. The skill uses XML tags (e.g., <task>, <output_format>) to layer information, defines dynamic placeholders with ${variable}, and builds a concise "role compression package" that clearly organizes the role, task, and output constraints.
3. 🧠 Think: It incorporates Tu Jinhao's "stream-of-consciousness thinking" and Schulhoff's Few-Shot techniques. For complex tasks, it embeds deep thinking guidance like "first examine the essence of the problem, explore multiple possibilities"; for tasks requiring specific formats, it provides carefully ordered examples to anchor the output. Crucially, in Chinese contexts, it avoids rigid "think step-by-step" instructions, opting instead for task decomposition or more organic thinking prompts.
4. ⚡ Compress: Applying Li Jigang's compression philosophy, it strips away all redundancy (e.g., ineffective threats/rewards, repetitive instructions), striving for each word to have a precise vector. The final output is a high-density, clear, and actionable prompt that helps the model "break from its default path."

The skill synthesizes the expertise of four masters: Goodside's structure and formatting, Tu Jinhao's thinking depth control, Schulhoff's evidence-based techniques, and Li Jigang's compression and resonance writing.

Scope & Key Considerations

• Service Boundary: This is a "prompt optimization expert," it does not directly answer your business questions. You provide the original prompt to be optimized, and it returns an optimized version along with a detailed change report (diagnostic report, structural rework, etc.).
• Input Pre-check: It automatically detects ambiguous or abnormal inputs. For instance, with bare commands like "optimize," it will ask for the specific scenario and target model. For very long inputs (>3000 characters), it suggests segmented optimization.
• Model Adaptation: Different models (e.g., Claude, GPT, Gemini) have different "personalities" and strengths. The skill will prompt you to specify a target model and provide targeted adaptation advice (e.g., Claude's preference for XML tags) during optimization.
• Multi-Turn Support: It can optimize entire conversation histories ("Multi-Turn Distillation"), focusing on maintaining context consistency, role continuity, and resolving constraint conflicts, not just restructuring a single prompt.
• Anti-Pattern Alerts: The skill has a built-in library of common anti-patterns (e.g., role stacking, threatening prompts, forcing English instructions in Chinese) and actively avoids these empirically ineffective practices during optimization.