Prompt Engineering

Folder: .github/skills/tsh-engineering-prompts/
Used by: Prompt Engineer, Software Engineer, Architect, Code Reviewer

Technology-agnostic patterns for designing, optimizing, and securing LLM application prompts. Covers system prompts, user prompt templates, RAG templates, agent instructions, and classification/extraction prompts. NOT for Copilot customization — use tsh-creating-prompts for that.

Prompt Structure Patterns

Reliable prompt structure separates concerns into distinct sections: system prompt (persona + rules), context (retrieved docs, session state), user input (the variable part), and output format (expected response shape). Patterns include role-based structure, few-shot prompting, chain-of-thought reasoning, and delimiter separation.

Optimization Techniques

Technique	Purpose
Clarity and specificity	Replace vague instructions with exact behavior, constraints, and output format
Constraint specification	Explicit MUST / MUST NOT rules for model behavior
Output format control	JSON schemas, structured output modes, explicit format instructions
Token efficiency	Remove filler, use tables over paragraphs, cache static content in system prompts
Negative prompting	Tell the model what NOT to do for better compliance
Temperature guidance	Match temperature to task type (0.0–0.2 for extraction, 0.7–1.0 for creative)

Security Patterns

Three-layer defense against prompt injection:

1. Delimiter separation — Always separate user input from instructions with clear delimiters.
2. Input sanitization — Strip delimiter characters, truncate, flag instruction-like patterns.
3. Output validation — Parse into typed models, validate schemas, reject unexpected content.

Also covers jailbreak resistance, secrets management (never hardcode in templates), and PII handling.

Prompt Templates

Reusable templates for common patterns:

• RAG — Context injection with grounding instructions and citation requirements
• Agent tool-calling — Tool definitions with usage rules and error handling
• Classification / Extraction — Category selection with confidence scoring
• Summarization — Length-controlled summaries with preservation rules
• Evaluation / Scoring — Multi-criteria scoring with justification

Evaluation Approaches

• A/B testing — Fix one variable, use the same test set, measure objectively
• Metric-based comparison — Format compliance, factual accuracy, consistency, token efficiency, hallucination rate
• Edge case testing — Empty input, long input, injection attempts, unexpected language, special characters
• Version control — Store prompts as named versioned constants, tag versions, log which version generated each response

Anti-Patterns

Anti-Pattern	Instead Do
Vague instructions ("be helpful")	Specify exact behavior and output format
User input in system prompt	Separate with delimiters in dedicated section
No output format specification	Define explicit schema or structure
Prompt-only validation	Parse into typed models, validate schemas
Hardcoded secrets in templates	Use environment variables or secret managers
Testing only happy paths	Include adversarial, empty, long, and multilingual inputs

Connected Skills

• tsh-creating-prompts — For Copilot .prompt.md files (different domain, complementary).
• tsh-code-reviewing — For reviewing prompt code quality alongside application code.
• tsh-architecture-designing — For prompt strategy decisions as part of system architecture.