| name | ai-prompt-writer |
| description | Design, review, and optimize secure AI prompts using advanced prompt engineering patterns, safety frameworks, and injection mitigation strategies. |
| license | MIT |
Skill: AI Prompt Writer
Design, review, and optimize secure AI prompts using advanced prompt engineering patterns, safety frameworks, and injection mitigation strategies.
WHEN TO USE
- Engineering system prompts, LLM templates, or Copilot instructions.
- Auditing existing prompts for injection vulnerabilities or data leakage risks.
- Defining automated prompt evaluation metrics or safety checklists.
WHEN NOT TO USE
- When writing strict procedural code (like Python or Bash scripts) where natural language ambiguity is unacceptable.
- For managing low-level hardware or infrastructure where strict, deterministic APIs should be used instead of LLMs.
- When generating general project documentation without targeting an AI consumer (use
docs-writer instead).
Core Process
- Contextualize: Define explicit tasks, target audience, and precise constraints (format, length, tone).
- Pattern Selection: Apply Zero-Shot for simple tasks, Few-Shot for formats, Chain-of-Thought for reasoning, and Role Prompting for specialized contexts.
- Harden: Sanitize all dynamic inputs to prevent prompt injection and data leakage.
- Validate: Red-team against bias, harmful outputs, and edge cases.
Core Principles
- Clarity over Verbosity: Issue concise, unambiguous directives. Drop conversational filler.
- Strict Constraints: Dictate exact output structures (JSON, markdown) and scope boundaries.
- Secure Construction: Never directly interpolate untrusted input. Use parameterization or robust sanitization (e.g.,
Translate this text: [SANITIZED_USER_INPUT]).
- Data Minimization: Exclude sensitive, personal, or proprietary data from prompts.
- Neutrality: Enforce inclusive language; eliminate demographic or contextual assumptions.
Commands / Usage Patterns
- Role Setup:
You are a <role> with <years> experience in <domain>. Review <target> and identify <issues>.
- Chain-of-Thought:
Solve this step-by-step: 1. <step 1> ...
- Few-Shot Validation:
Input: <example_in> \n Output: <example_out>
- Injection Defense: Replace
`Text: ${userInput}` with `Text: ${sanitizeInput(userInput)}`.
Best Practices
- Standardize inputs using clear delimiters (e.g.,
""", ###, <tag>).
- Implement content moderation pre/post-processing for user-facing AI pipelines.
- Establish A/B testing and automated evaluation for continuous prompt refinement.
- Document prompt purpose, scope, assumptions, and limitations.
Common Pitfalls
- Too Vague: "Make the code better" → Specify exactly what "better" means.
- Too Restrictive: Over-constraining can prevent useful optimizations.
- Missing Context: Include relevant domain knowledge and terminology.
- No Examples: Concrete examples guide an LLM better than abstract descriptions.
- Ignoring Artifacts: Don't refine prompts based on error feedback.
What to Avoid
- Unbounded Inputs: Permitting raw user input to override system instructions.
- Overfitting: Hardcoding constraints tied to specific training data examples.
- Echoing Secrets: Designing prompts that inadvertently reflect sensitive inputs.
References
