gemini-ai-prompt-engineering

krasimir-hristov/gemini-ai-prompt-engineering

AI & ML

About

This skill governs all AI-related logic within the TalentFilter platform, specifically focusing on the AIService in the backend.

Model Versioning: Always use the most capable version (gemini-2.5-flash).
Safety Settings: Configure AI safety settings via the google-genai SDK to ensure the AI doesn't refuse valid recruitment/screening tasks while maintaining professional boundaries.
Async Execution: AI calls are high-latency; always use await and ensure they are proxied through the FastAPI backend to avoid blocking the main thread.

Strict Schema: Every AI call must use a Pydantic model to define its response structure.
JSON Request: Set the response_mime_type to application/json in the Gemini config.
Validation: Always wrap the AI parsing logic in a try-except block. If parsing fails, implement a retry mechanism or return a structured error response.
Language Awareness: Every prompt must explicitly specify the output language based on the user's locale.
Dynamic Scope: Avoid hardcoding question counts. Instruct the AI to determine the optimal depth (e.g., "Generate as many questions as needed to thoroughly vet this specific role").

System Instructions: Clearly define the AI's persona as an "Expert Technical Recruiter" or "Bias-Free Interviewer".
Chain-of-Thought: For complex tasks like grading interviews, use prompts that encourage step-by-step reasoning within the JSON output (e.g., a reasoning field before the final_score).
Context Injection:
- Job Analysis: Inject raw job description text + specific extraction goals.
- Suggest Question: Inject job description + existing questions (as JSON array) to avoid duplicates. Prompt must explicitly instruct: "Generate ONE new question that is NOT similar to existing ones."
- Generate Answer: Inject job description + specific question text. Prompt should ask for an "ideal answer" that reflects the company's requirements based on the description.
- Grading: Inject Question context + "Ideal Answer" + "Candidate Answer" + "Scoring Rubric".

Separation: Never hardcode long prompts in Python files.
Storage: Store prompt templates in backend/app/core/prompts/ as .txt or .md files.
Template Engine: Use simple string formatting or Jinja2 to inject variables into prompts.

Token Optimization: Be concise in prompts. Avoid redundant instructions.
Caching: If similar requests are expected (e.g., re-analyzing the same job description), check for existing results in Supabase before calling the AI.

SKILL.md

This skill governs all AI-related logic within the TalentFilter platform, specifically focusing on the AIService in the backend.

Model Versioning: Always use the most capable version (gemini-2.5-flash).
Safety Settings: Configure AI safety settings via the google-genai SDK to ensure the AI doesn't refuse valid recruitment/screening tasks while maintaining professional boundaries.
Async Execution: AI calls are high-latency; always use await and ensure they are proxied through the FastAPI backend to avoid blocking the main thread.

Strict Schema: Every AI call must use a Pydantic model to define its response structure.
JSON Request: Set the response_mime_type to application/json in the Gemini config.
Validation: Always wrap the AI parsing logic in a try-except block. If parsing fails, implement a retry mechanism or return a structured error response.
Language Awareness: Every prompt must explicitly specify the output language based on the user's locale.
Dynamic Scope: Avoid hardcoding question counts. Instruct the AI to determine the optimal depth (e.g., "Generate as many questions as needed to thoroughly vet this specific role").

System Instructions: Clearly define the AI's persona as an "Expert Technical Recruiter" or "Bias-Free Interviewer".
Chain-of-Thought: For complex tasks like grading interviews, use prompts that encourage step-by-step reasoning within the JSON output (e.g., a reasoning field before the final_score).
Context Injection:
- Job Analysis: Inject raw job description text + specific extraction goals.
- Suggest Question: Inject job description + existing questions (as JSON array) to avoid duplicates. Prompt must explicitly instruct: "Generate ONE new question that is NOT similar to existing ones."
- Generate Answer: Inject job description + specific question text. Prompt should ask for an "ideal answer" that reflects the company's requirements based on the description.
- Grading: Inject Question context + "Ideal Answer" + "Candidate Answer" + "Scoring Rubric".

Separation: Never hardcode long prompts in Python files.
Storage: Store prompt templates in backend/app/core/prompts/ as .txt or .md files.
Template Engine: Use simple string formatting or Jinja2 to inject variables into prompts.

Token Optimization: Be concise in prompts. Avoid redundant instructions.
Caching: If similar requests are expected (e.g., re-analyzing the same job description), check for existing results in Supabase before calling the AI.