cv-tailoring

name

cv-tailoring

description

Use when implementing or modifying the CV tailoring pipeline — selecting the right CV version, reordering project/skill blocks for a specific job, generating dynamic cover letters, or updating the tailoring prompt contract. Triggers on: "CV seçimi", "CV tailoring", "özgeçmiş optimizasyonu", "cover letter", "immutable block selection", "CV versiyonu", "keyword injection", "tailored summary", "proje sıralama".

CV Tailoring Skill

Use this skill when

• Selecting the optimal CV version for a specific job listing
• Reordering skill/project blocks without fabricating new content
• Generating a dynamic cover letter or "brief" PDF
• Updating the tailoring prompt contract or JSON schema
• Debugging why the wrong CV version was selected

Do not use this skill when

• General LLM prompt engineering unrelated to CV content
• Managing CV files on disk/S3 (that's infrastructure, not this skill)

Core Principle: Immutable Block Selection

The agent NEVER generates false experience or skills. It only:

1. Selects which existing blocks to include (from Master Data)
2. Reorders blocks by relevance to the JD
3. Rewrites only the Summary (1-2 sentences, with JD keywords)

Instructions

Step 1: Select CV Version

async def select_cv_version(job: JobListing, cv_docs: list[CVDocument]) -> CVDocument:
    """
    Use LLM to pick the best CV version from available options.
    cv_docs.metadata contains keywords extracted from each CV PDF.
    """
    prompt = f"""
    Job requires: {job.detected_stack}
    Available CV versions: {[{'id': cv.id, 'name': cv.version_name, 'keywords': cv.metadata['keywords']} for cv in cv_docs]}
    Return ONLY the id of the best matching CV version as JSON: {{"selected_cv_id": "..."}}
    """
    # Use call_with_fallback('cv_tailoring', prompt)

Step 2: Tailor Block Order (Prompt Contract)

System prompt:

Görevin, adayın ham veri setinden iş ilanına en uygun olanları seçip sıralamak.
KURAL: Yeni veri üretme, var olan cümleleri değiştirme. Sadece ID'leri kullanarak sıralama yap.

Input payload:

{
  "master_skills": [{"id": 10, "name": "FastAPI", "category": "backend"}, ...],
  "master_projects": [{"id": 1, "title": "OpenReef", "tech_stack": ["FastAPI", "LLM"]}, ...],
  "job_description": "..."
}

Output schema (validate with Pydantic):

class TailoringOutput(BaseModel):
    tailored_summary: str = Field(max_length=300)
    selected_project_ids: list[int] = Field(min_items=1, max_items=5)
    top_skill_ids: list[int] = Field(min_items=3, max_items=10)
    keyword_injections: list[str] = Field(max_items=5)

Step 3: Generate Cover Letter / Brief PDF

# Use reportlab or fpdf2
from fpdf import FPDF

def generate_cover_letter(tailoring: TailoringOutput, job: JobListing, profile: UserProfile) -> str:
    pdf = FPDF()
    pdf.add_page()
    # ... populate with tailored_summary, selected projects, keyword_injections
    output_path = f"/storage/cover_letters/{job.id}_{datetime.now().strftime('%Y%m%d')}.pdf"
    pdf.output(output_path)
    # Save path to applications.application_data['cover_letter_path']
    return output_path

Step 4: Archive

After generation, update applications table:

UPDATE applications SET 
    cv_id = $1,
    application_data = application_data || '{"cover_letter_path": "$2", "tailored_keywords": $3}'
WHERE id = $4;

Model

Use claude-sonnet-4 as primary (quality matters for CV language). Fallback: gpt-4o-mini → groq/llama-3-70b → ollama/llama3:8b (See rate-limit-fallback skill for chain implementation)

Performance Tracking

Query cv_performance_stats VIEW regularly to see which CV version has the highest conversion_rate. Switch default version if another outperforms by >5%.