Prompt X-Ray Skills Router

Purpose

Use this root skill as a router for the Prompt X-Ray skill pack.

The goal is to choose the right prompt-engineering workflow and apply it consistently.

Do not treat this repository as a list of prompt templates. Treat it as a set of workflows for designing reliable LLM behavior.

Core principle

A prompt is a behavioral specification for an AI system.

A good prompt should usually define:

1. Goal
2. Context
3. Role
4. Inputs
5. Constraints
6. Source of truth
7. Decision rules
8. Output format
9. Uncertainty behavior
10. Quality checks

If a prompt cannot be tested, it is not production-ready.

Available workflows

Use these internal skills depending on the task.

01-prompt-brief-builder

Use when the user has a vague idea and needs a structured prompt brief before writing the final prompt.

Typical triggers:

• "Help me write a prompt"
• "I need an AI assistant that..."
• "Turn this idea into a prompt"
• "I don't know how to explain this to an LLM"

Output:

• prompt brief;
• inputs;
• source of truth;
• constraints;
• output format;
• test cases.

02-prompt-auditor

Use when the user provides an existing prompt and wants to improve, harden, debug, or make it more reliable.

Typical triggers:

• "Improve this prompt"
• "Audit this prompt"
• "Make this prompt production-ready"
• "Why is this prompt unreliable?"

Output:

• prompt audit report;
• main problems;
• missing information;
• risk areas;
• improved prompt;
• test cases.

03-system-prompt-architect

Use when the user needs a system prompt for an assistant, chatbot, internal tool, or workflow agent.

Typical triggers:

• "Create a system prompt"
• "Write instructions for an AI assistant"
• "Design assistant behavior"
• "Make a prompt for a support bot"

Output:

• system prompt;
• mission;
• scope;
• source-of-truth policy;
• workflow;
• uncertainty rules;
• output contract.

04-agent-instruction-writer

Use when the user needs instructions for coding agents or project agents.

Typical triggers:

• "Write AGENTS.md"
• "Create CLAUDE.md"
• "Make Cursor rules"
• "Prepare Codex instructions"
• "Tell the coding agent how to work in this repo"

Output:

• repository-level agent instructions;
• architecture map;
• allowed/forbidden behavior;
• command policy;
• testing policy;
• reporting format.

05-rag-answer-policy

Use when the prompt is for a knowledge-base assistant or RAG system.

Typical triggers:

• "Make a RAG prompt"
• "The bot should answer only from docs"
• "Define no-answer behavior"
• "Prevent unsupported answers"

Output:

• source priority;
• citation rules;
• insufficient-context behavior;
• conflicting-source behavior;
• escalation rules.

06-structured-output-designer

Use when the LLM output must be parsed, stored, validated, or consumed by software.

Typical triggers:

• "Return JSON"
• "Create a schema"
• "Extract fields"
• "Make CRM-ready output"
• "Design structured output"

Output:

• output schema;
• field definitions;
• validation rules;
• valid example;
• insufficient-context example.

07-prompt-test-generator

Use when the user needs test cases or quality checks for a prompt.

Typical triggers:

• "Test this prompt"
• "Generate evals"
• "Create edge cases"
• "How do I check if this prompt works?"

Output:

• quality-check suite;
• typical cases;
• edge cases;
• incomplete-context cases;
• ambiguous cases;
• format-stability cases;
• review criteria.

08-prompt-debugger

Use when the prompt is already producing bad outputs and the user wants to diagnose the cause.

Typical triggers:

• "The model ignores my format"
• "This output is too generic"
• "Why is this answer bad?"
• "Debug this prompt"

Output:

• debug report;
• expected vs actual behavior;
• likely causes;
• smallest useful fix;
• revised prompt;
• confirmation test.

09-prompt-to-product-spec

Use when a prompt or assistant idea should become a product specification or MVP task.

Typical triggers:

• "Turn this AI idea into a spec"
• "Create a product spec for this assistant"
• "Convert this prompt into a product feature"
• "Prepare this for a developer"

Output:

• product specification;
• users and roles;
• AI responsibilities;
• human responsibilities;
• MVP scope;
• data sources;
• acceptance criteria;
• risks.

Routing procedure

When the user asks for prompt-engineering help:

1. Identify the task type.
2. Select the most relevant internal skill.
3. If the request is vague, start with 01-prompt-brief-builder.
4. If the user provides an existing prompt, use 02-prompt-auditor.
5. If the output must be consumed by software, include 06-structured-output-designer.
6. If the prompt will be reused, include 07-prompt-test-generator.
7. If the user wants implementation planning, use 09-prompt-to-product-spec.
8. Produce a concrete artifact, not generic advice.

Common chains

Vague prompt idea

01-prompt-brief-builder -> 02-prompt-auditor -> 07-prompt-test-generator

Assistant system prompt

01-prompt-brief-builder -> 03-system-prompt-architect -> 02-prompt-auditor

Coding agent instructions

04-agent-instruction-writer -> 02-prompt-auditor -> 07-prompt-test-generator

RAG assistant

05-rag-answer-policy -> 03-system-prompt-architect -> 07-prompt-test-generator

Structured extraction/API output

06-structured-output-designer -> 02-prompt-auditor -> 07-prompt-test-generator

Broken prompt output

08-prompt-debugger -> 02-prompt-auditor -> 07-prompt-test-generator

Prompt idea to MVP

01-prompt-brief-builder -> 09-prompt-to-product-spec

Output rules

When applying a workflow:

• produce the requested artifact directly;
• avoid generic prompt tips;
• mark assumptions clearly;
• separate stable instructions from variable inputs;
• define missing-information behavior;
• include output format when relevant;
• include quality checks for reusable prompts;
• prefer practical examples over theory.

Default behavior

If the user does not specify a format, return:

1. Diagnosis or brief summary
2. Main artifact
3. Quality checklist
4. Suggested next step

Boundaries

This skill routes prompt-engineering workflows. It does not guarantee identical model behavior across platforms. Always test important prompts with realistic inputs before production use.