Skip to content

PROMPTS.md

Latest commit

 

History

History
54 lines (37 loc) · 2.95 KB

PROMPTS.md

File metadata and controls

54 lines (37 loc) · 2.95 KB

PROMPTS & AI Methodology

This document outlines the system instructions (prompts) used by Gemini to generate, score, and optimize resumes. It also explains the rationale behind the scoring algorithms.

1. Resume Generation & Optimization

File: src/models/prompts.py Trigger: /api/generate

The generation prompt instructs Gemini to act as an "Elite ATS Resume Strategist".

Key Instructions:

  • Non-Negotiable Constraints: Never fabricate data (companies, dates, degrees). If data is missing (e.g. skills), infer it from experience, but do not invent experiences.
  • Structure: Enforces specific ATS-friendly sections: Summary, Skills, Experience, Projects, Education, Certifications.
  • Formatting: Enforces "MMM YYYY" date format and single-line categorised skill lists to save space.
  • Hyperlink Preservation: Explicit instruction to preserve GitHub/LinkedIn URLs, which AI models often strip by mistake.
  • Keyword integration: Instructions to naturally weave JD keywords into bullet points.

2. ATS Scoring Engine (Hybrid Approach)

File: src/services/ats_scorer.py & src/models/prompts.py Trigger: /api/score (or part of /api/generate)

We use a Hybrid Scoring System that combines deterministic algorithmic checks with AI semantic evaluation.

Algorithm (Code-based):

  • Keyword Match (30%): Calculates the percentage of JD keywords present in the resume plain text.
  • Section Completeness (20%): Checks for the existence of required headers (Summary, Education, Experience, etc.).
  • Formatting (10%): Checks for parsing blockers like tables or columns (though our generator guarantees clean output).

AI Evaluation (Gemini-based):

  • Role Alignment (25%): AI evaluates how well the narrative of the resume fits the target role seniority and domain.
  • Content Quality (15%): AI scores the usage of action verbs ("Engineered" vs "Worked on") and quantified metrics (e.g., "Reduced latency by 20%").

Rationale: purely AI scoring can be hallucinated or inconsistent. Pure algorithmic scoring misses context (e.g., "React.js" context vs just the word "React"). Combining them gives the most accurate "human recruiter + machine ATS" simulation.

3. Section Enhancement

File: src/models/prompts.py -> SECTION_ENHANCEMENT_PROMPT Trigger: /api/enhance

This prompt is a focused "micro-editor". It takes a single string (e.g., a summary or a job description block) and rewrites it to be:

  1. Concise: Removes fluff.
  2. Impactful: Forces action-verb starts.
  3. Keyword-Rich: Injects keywords from the provided JD.

4. Chatbot Persona

File: src/models/prompts.py -> CHATBOT_SYSTEM_PROMPT Trigger: /api/chat

The chatbot is instructed to be a "Recruiter-style Career Advisor".

  • Tone: Direct, professional, no fluff "motivational speaker" style.
  • Capabilities: analyzing JD gaps, suggesting specific projects to build, explaining why a score is low.