prompt-optimize_skill

Overview

This skill transforms the assistant into Alpha-Prompt, an expert prompt engineer that collaborates with users to design, refine, and harden high-quality prompts and system instructions. It emphasizes conversational co-creation, practical architectures, and safety measures to deliver usable, production-ready prompts. The goal is to help you get reliable, reproducible AI behavior through clear structure and expert choices.

How this skill works

When activated by requests like "optimize prompt" or "improve system instruction", the skill runs a three‑phase workflow: diagnose requirements through targeted questions, co-build or critique prompt drafts, and deliver a final, copy‑ready prompt plus design rationale. It recommends cognitive architectures (e.g., Chain-of-Thought, Tree-of-Thought, ReAct) and output formats (XML/Markdown/JSON) appropriate to the task, and attaches safety guardrails where public or adversarial risk exists. The interaction is iterative and waits for user decisions at every key step.

When to use it

• You want to improve an existing prompt, system instruction, or AI role definition.
• You need a prompt designed for a specific application (summaries, creative generation, reasoning).
• You're building a public-facing or user-interactive agent and need safety constraints.
• You want to upgrade a simple request into a stronger architecture (CoT, ToT, ReAct).
• You have a high-quality prompt and want peer-review style, advanced refinements.

Best practices

• Start with clear persona, explicit goals, and hard constraints before drafting prompts.
• Choose architecture to match cognitive needs: CoT for complex reasoning, ToT+Self-Consistency for creativity.
• Prefer structured output (XML/JSON) when precision is required; use concise Markdown for readability.
• Add instruction fencing and refusal policies for public-facing roles to reduce injection risk.
• Iterate with the user: ask clarifying questions and wait for decisions at key tradeoffs.

Example use cases

• Optimize a marketing prompt: propose a two-step generate-then-evaluate pipeline (ToT + Self-Consistency).
• Design a customer-support persona with explicit safety constraints and graceful refusal behavior.
• Convert meeting notes into structured XML summaries with defined fields for actions and owners.
• Peer-review a polished prompt: praise strengths, propose deep architectural upgrades, then ask for direction.
• Create a reasoning prompt that uses Chain-of-Thought and step-back prompting for deeper insight.