prompt-architect_skill

Overview

This skill designs and refines prompts, system instructions, and principle files for production use. It transforms verbose prompts into structured, token-efficient artifacts and shapes CLAUDE.md-style principle files and agent personas. Use it to create, compress, optimize, or refactor prompts while preserving domain depth and intended role.

How this skill works

The skill diagnoses the input type (create, enhance, or ask) then applies targeted transformations: preserve role and domain knowledge, replace repetitions with domain terms when safe, add reasoning patterns (CoT, ReAct, few-shot) when needed, and restructure principle files into soul/mental-models/tensions. It outputs only the prompt or principle file in a parsable format chosen to match the user's needs.

When to use it

• Creating a new system prompt, agent role, or persona for an autonomous agent
• Improving CLAUDE.md or principle files to capture identity and thinking patterns
• Compressing verbose prompts while preserving expert-level detail
• Adding reasoning patterns (chain-of-thought, ReAct, few-shot) to increase reliability
• Optimizing token usage for cost-constrained deployments
• Refactoring workflow prompts and rule sets into maintainable sections

Best practices

• Preserve domain knowledge depth—never remove detail without a semantically equivalent term
• Keep the role when present; output must include role if input had one
• Prefer domain taxonomy or specialized terms in place of enumerations when equivalent
• Apply reasoning techniques only when the task demands complex judgment
• Ask one focused question if the intent or scope is unclear

Example use cases

• Turn a long product-spec prompt into a compact system prompt that preserves legal and compliance details
• Convert a loose persona sketch into a CLAUDE.md soul section with thinking_style and tensions
• Add CoT scaffolding and few-shot examples to a prompt used for high-stakes reasoning
• Compress customer-support workflows using taxonomy references (e.g., HTTP, error classes) to replace examples
• Refactor an agent prompt to include mental models and explicit escalation boundaries