Prompt Writer

Convert vague instructions, rough briefs, or messy user requests into structured, high-performance prompts for LLMs. Use whenever the user asks to "rewrite this prompt," "improve this prompt," "fix this prompt," "make this prompt better," "turn this into a prompt," "make this clearer for an LLM," or anything else involving prompt engineering, prompt optimization, or converting natural language instructions into LLM-ready prompts. Trigger this skill even when the user just pastes a rough request and says "make this better" without naming it as a prompt task. The skill applies to system prompts, user prompts, agent instructions, and any other LLM input.

1. Identify the primary task, constraints, and implicit context from the rough user request.
2. Assign a specific persona that activates the right region of the training distribution (e.g. 'senior frontend engineer specializing in Bloomberg Terminal').
3. Structure the task sequentially, with planning or research steps where needed, and add verification loops for quality assurance.
4. Encode every constraint concretely and falsifiably (e.g. 'wasted whitespace is a failure state' instead of 'be data dense'), then suppress default LLM behaviors (hedging, over-explaining, presenting options).
5. Specify the exact output format: sections, length targets, headers, bullets, code blocks, and what to exclude. Output: a structured, high-performance prompt ready to execute against any LLM.

Convert vague instructions, rough briefs, or messy user requests into structured, high-performance prompts for LLMs.

The methodology traces to three mechanical principles of how LLMs generate text: (1) early tokens condition all tokens that follow, so front-load critical instructions; (2) specificity activates narrower regions of training data, so use concrete references ('Bloomberg Terminal') over abstract descriptions; (3) suppression of default behaviors matters as much as positive instruction, so explicitly close off hedging, over-explaining, and presenting options. Standard prompt architecture: persona (1-3 sentences) → critical context → task statement → sequential structure → constraints and rules (falsifiable) → suppression instructions → examples (if needed) → output format specification → verification step (if applicable). Use why-explanations sparingly to help the model decide edge cases consistently.

Do not apply this skill when the user is asking a question, having a conversation, or making a request that you will execute yourself. The skill is for crafting input to a model, not for being responsive to a person. Over-specification on small tasks wastes tokens and dilutes signal; under-specification on large tasks produces generic output. Match prompt length to task complexity. Avoid vague positive instructions ('be good,' 'make it nice'), buried critical instructions, and conflicting instructions within the same prompt.