Python + Streamlit + Claude API | Built on Structured Prompt Engineering
Paste a JD and resume — get a match score, inline gap analysis, and specific improvement actions in seconds
Job seekers get rejected by ATS systems without knowing why. Recruiters spend 5-10 minutes per resume doing a comparison that could take 5 seconds. Companies write vague JDs that attract the wrong candidates.
This tool fixes all three in one place.
The same job description was tested against a weak resume and a strong resume.
Weak resume — 28/100 Poor Match:
Strong resume — 92/100 Strong Match:
64 points difference. Same JD. Same candidate background. Different language, structure, and keyword alignment. The system shows exactly what separates the two.
Tested on multiple real-world JDs and resumes across different roles and experience levels.
- Job seekers who want to know exactly why their resume is not getting callbacks
- Recruiters who want faster, consistent resume screening with a structured score
- Hiring teams who want ATS-optimized job descriptions without copy-pasting old ones
Most people think they can paste a resume into ChatGPT and get the same result. They cannot.
| Feature | ChatGPT | This Tool |
|---|---|---|
| Consistent scoring structure | No — output varies every run | Yes — weighted dimension table |
| Hallucination prevention | No — invents missing skills | Yes — strict rules enforced in prompt |
| Inline gap per skill | No — generic feedback | Yes — found in which section OR where to add |
| Before/after examples | Rarely | Yes — every action has one |
| Downloadable PDF report | No | Yes |
| Bias-free JD generation | No | Yes — banned words list enforced |
The difference is not the model. It is how the prompt is engineered.
Paste a JD and a resume. Get:
- Match score 0-100 with colour coding (red/yellow/green)
- Weighted dimension breakdown with reasoning
- Skill match chips — green for matched, red for missing critical, yellow for missing preferred
- Inline gap analysis — per skill found or missing, with where to add it
- Top 5 improvement actions with before/after examples
- One resume section fully rewritten to match the JD
- Honest assessment — direct verdict, no softening
- Downloadable PDF report
Not just "you are missing Power BI." It says "Power BI is missing — add it to your Skills section and mention it in your Analytics Engineer role." That specificity is what makes the output actionable.
Fill in 7 fields. Get a structured, bias-free, ATS-optimized job description.
Every JD includes a built-in Quick Check — bias scan, top 5 ATS keywords embedded, and missing information flagged. The generated JD passes directly into the Resume Analyzer with one click.
Full analysis exported as a formatted PDF — match score, breakdown, gap analysis, improvement actions, section rewrite, and honest assessment. Useful for candidates to track progress or recruiters to attach to candidate records.
This project shows real prompt engineering — not theory. Three prompt versions for each module, each with documented failures and specific fixes.
V1 — Basic:
Compare this resume with this JD. Give a score out of 100.
Failed: single number, no explanation, inconsistent, hallucinated skills not in either document.
V2 — Structured:
You are a recruiter. Provide: score, matched skills, missing skills, 3 suggestions.
Improved: structure added. Still failed: generic suggestions, no dimension breakdown, model still invented skills.
V3 — Final: Added strict hallucination rules, weighted dimension table, Critical vs Preferred skill split, inline gap per skill, before/after required in every action, direct honest assessment forced.
V1 failed: Used "rockstar" and "ninja", 15 unordered requirements, different structure every run.
V2 improved: Persona, sections, professional tone. Still failed: no bias prevention, no bullet count.
V3 fixed: Banned words list, exactly 6 bullets enforced, Required max 6 / Preferred max 4, built-in self-evaluation.
Full prompt files are in the prompts/ folder — V1, V2, V3 for both modules.
| Scenario | Response |
|---|---|
| Empty JD | Error: Please paste a job description |
| Empty resume | Error: Please paste your resume |
| Resume under 30 words | Error: Resume too short |
| Resume over 3000 words | Warning: Consider trimming to 2 pages |
| Non-English text | Notice: Analysis may be less accurate |
| Vague JD | Notice: Analysis focuses on experience alignment |
git clone https://github.com/analytics-ak/ai-hiring-workflow-engine.git
cd ai-hiring-workflow-engine
pip install -r requirements.txt
# Add ANTHROPIC_API_KEY to .env
python -m streamlit run app.pyTest immediately using the sample files in sample_data/:
sample_jd_good.txt+sample_resume_weak.txt→ expect ~28/100sample_jd_good.txt+sample_resume_strong.txt→ expect ~92/100
ai-hiring-workflow-engine/
├── app.py <- Streamlit entry point
├── modules/
│ ├── jd_generator.py <- JD Generator + V3 prompt
│ └── resume_analyzer.py <- Resume Analyzer + V3 prompt
├── prompts/
│ ├── jd_v1.txt / v2.txt / v3.txt
│ └── resume_v1.txt / v2.txt / v3.txt
├── utils/
│ ├── report_generator.py <- PDF generation
│ └── scorer.py <- JD quality scoring
├── sample_data/ <- 4 demo files for testing
├── screenshots/ <- app screenshots
├── .env <- API key — gitignored
├── requirements.txt
└── README.md
Prompt Engineering — V1 to V3 evolution with documented failures, measurable improvements, real output comparisons. Strict hallucination prevention. Structured output enforcement.
Streamlit Development — Multi-page app, session state for cross-module data passing, color-coded UI, progress bars, download buttons.
Product Thinking — Two modules connected by workflow. JD Generator output feeds directly into Resume Analyzer. Every feature designed around how users actually work.
PDF Generation — In-memory PDF with ReportLab, color-coded score card, structured sections, served as download without saving to disk.
| Tool | Purpose |
|---|---|
| Python | Core language |
| Streamlit | Web UI |
| Anthropic SDK — claude-sonnet-4-5 | AI analysis and generation |
| ReportLab | PDF report generation |
| python-dotenv | API key management |
- Scoring may vary slightly between runs — it is a guide, not a definitive pass/fail
- Analysis quality depends on how clearly the JD and resume are written
- Uses rule-based prompt engineering — does not learn from previous analyses
- Module 3 — Resume Optimizer (full resume rewrite for target JD)
- Module 4 — Candidate Communication (rejection and shortlist emails)
- Module 5 — HR Tools (performance reviews, escalation reports)
Ashish Kumar Dongre