DIY Recruitment — AI CV Screening Agent

Section 01 · Architecture

Four stages, one AI agent, zero manual sorting.

Each stage feeds the next. Claude runs extraction, scoring and ranking automatically — you manage the inputs and review the shortlist. This is a human-in-the-loop workflow: the agent never contacts candidates or makes a hiring decision.

1

Ingest

Pull CVs from Indeed.ae folder + unzip the WhatsApp archive. Normalise PDF, DOCX and image into one text pool.

2

Extract

AI reads every CV. Extracts name, phone, email, nationality, UAE experience, languages and retail/sales background.

3

Score

Claude scores each candidate against your weighted job rubric on a 0–100 scale and ranks the full list.

4

Output

Generates a searchable HTML dashboard — filter by score, language and experience. Top 10 flagged. Excel export.

Design principle. The AI does the heavy reading and scoring. You review the shortlist and make the hire call. Cancellations, offers and candidate contact stay in your hands — exactly where regulators (and good practice) want them.

Section 02 · Pipeline

The 4-step agent pipeline.

Each step is clearly split between what you do and what the agent does — with the actual prompts you can copy-paste into Claude Code.

1

Manual setup

Ingest — collect every CV into one folder.

Two sources feed the agent: Indeed.ae (1,000+ CVs from your employer dashboard) and 40 WhatsApp CVs sent by friends, agencies and walk-ins. This step merges them into a single input folder the agent will process.

Indeed.ae bulk download. Log into your employer account → Manage Applications → Select All → Download as ZIP. For anything over 50 CVs, use Claude Code's browser agent to loop the dashboard and save every file to ~/Downloads/indeed-cvs/ automatically.

WhatsApp export. On phone, open the chat → select files → Share → Export → save the ZIP to Downloads. On desktop, unzip into ~/Downloads/whatsapp-cvs/.

Claude Code · Browser agent # Task: Download all CVs from Indeed.ae employer dashboard browser.navigate("https://employer.indeed.com/jobs/applications") browser.wait_for("#applications-list") for each application_row: click("Download CV" button) save_file(f"~/Downloads/indeed-cvs/{candidate_name}.pdf") # Runs headlessly — downloads ~1,000 CVs in ~45 min

📁 ~/Downloads/indeed-cvs/ 📁 ~/Downloads/whatsapp-cvs/ ✓ Target ≈ 1,040 files Formats: PDF · DOCX · JPG · PNG

2

You provide · AI configures

Job description — paste your Promoter JD.

Drop the Indeed.ae job description into the agent. The JD becomes the scoring rubric. The agent extracts required skills, experience level, language requirements, nationality preferences, education minimum, and any deal-breaker criteria.

Agent system prompt """ You are an HR screening agent for a UAE retail company. Job Title: Promoter / Brand Ambassador Location: Dubai, UAE # From the job description, extract these scoring criteria: REQUIRED: UAE experience, retail/sales background, English fluency PREFERRED: Arabic speaker, visa status (visit/transferable), presentation standards, own transport DEAL-BREAKERS: No UAE experience, no English, no customer-facing background # Score each candidate 0–100 against the weighted rubric below. """

📋 Paste JD here 🤖 Auto-extracts criteria ⚖️ Builds scoring rubric

3

AI automated

Extract — the agent reads every CV.

Claude Code processes all 1,040 CVs one by one. For each file (PDF, DOCX or image), the agent extracts structured data into a master JSON. Images are OCR'd. Scanned PDFs are parsed. Badly formatted CVs are flagged, not skipped.

Extraction agent loop # For each CV file in ~/Downloads/all-cvs/: extracted = claude.extract(cv_file, schema={ "full_name": str, "email": str, "phone": str, "nationality": str, "visa_status": ["employment","visit","transferable","own"], "uae_experience_years": int, "retail_experience": bool, "languages": list[str], "last_employer": str, "education": str, "source": "indeed" | "whatsapp" }) master_json.append(extracted) # Output: candidates.json (~1,040 records)

📄 1,040 CVs → JSON 🔍 OCR for image CVs 📊 15 fields per CV ⚠ Unreadable CVs flagged

4

Final output

Score, rank & build the searchable dashboard.

The agent scores every candidate against the weighted rubric, ranks all 1,040 and writes a self-contained HTML file you can open in any browser. Searchable. Filterable. Sortable. Top 10 highlighted. Export to Excel in one click.

Scoring + output agent # Score each candidate using the weighted rubric def score_candidate(candidate, job_description): score = 0 score += weight("uae_experience", 25) # max 25pts score += weight("retail_sales_bg", 25) # max 25pts score += weight("english_fluency", 20) # max 20pts score += weight("arabic_speaker", 10) # max 10pts score += weight("visa_status", 10) # max 10pts score += weight("education", 5) # max 5pts score += weight("availability", 5) # max 5pts return score # 0–100 # Build HTML dashboard with search + filters generate_html(ranked_candidates, top_n=10)

🏆 Top 10 auto-selected ✅ HTML dashboard built 🔍 Search by name / skill / score 📊 Excel export button

Section 03 · Scoring rubric

The Promoter rubric — 100 points, transparent weights.

Paste your JD and the agent auto-populates these weights. You can adjust before running. Everything below is editable — the rubric is a starting frame, not a verdict.

Criterion	Weight	How AI scores it	Deal-breaker?
UAE work experience Retail, FMCG, promotions, events	25	2+ yrs = 25 · 1 yr = 15 · <1 yr = 5 · None = 0	Yes — 0 UAE exp auto-drops below rank 900
Retail / sales background Supermarket, mall, FMCG, brand	25	Promoter = 25 · Sales/retail = 18 · Customer-facing = 10 · Unrelated = 0	Soft — strongly preferred
English fluency CV written in English used as proxy	20	Fluent EN = 20 · Mixed = 12 · Non-English only = 5	Near-breaker for Dubai retail
Arabic speaker Native or professional	10	Arabic listed = 10 · Not listed = 0	No — bonus only
Visa / availability Visit, transferable, freelance	10	Visit = 10 · Transferable = 8 · Own = 6 · Requires new = 2	No — affects speed to hire
Education Min secondary school	5	Degree = 5 · Diploma = 4 · Secondary = 3 · Not stated = 1	No
Immediate availability Can start within 2 weeks	5	Immediate = 5 · 1-month notice = 3 · Not stated = 2	No

Total = 100. ≥ 75 = auto-shortlist · 60–74 = review pool · < 60 = archived but accessible. Top 10 by score = your recommended interview list.

Section 04 · Dashboard output

A single HTML file. No server, no login, no cloud.

The agent writes one self-contained file. Open it in any browser. Search, filter, sort 1,040 candidates instantly. Top 10 highlighted. Print or export to Excel.

cv-screening-results.html — 1,040 candidates loaded

Promoter Candidate Dashboard 1,040 CVs · Top 10 flagged 📊 Export Excel

#	Candidate	Nationality	UAE exp	Languages	Visa	Score	Status
★1	Sarah Mohammed Al-Farsi Lulu Hypermarket · 3 yrs · Indeed	UAE National	3 yrs	EN · AR	Own	94	Shortlist
★2	Priya Rajan Nair Carrefour UAE · 2.5 yrs · Indeed	Indian	2.5 yrs	EN · HI	Transfer	91	Shortlist
★3	Maria Santos Cruz Spinneys Dubai · 2 yrs · WhatsApp	Filipino	2 yrs	EN · FIL	Visit	88	Shortlist
4	Fatima Al-Qasimi Landmark Group · 1.5 yrs · Indeed	Egyptian	1.5 yrs	EN · AR	Transfer	82	Shortlist
5	Ravi Kumar Sharma Noon.com · 1 yr · Indeed	Indian	1 yr	EN · HI · UR	Own	79	Shortlist
6	Aisha Benali Morocco — no UAE exp · Indeed	Moroccan	0 yrs	EN · FR · AR	Visit	48	Archive

Showing 6 of 1,040 · Sorted by Score ▼ Click any row → full CV viewer

Top 10 — final shortlist preview

1

Sarah Mohammed Al-Farsi
Lulu Hypermarket · UAE National

94

3 yrs

✓

Shortlist

2

Priya Rajan Nair
Carrefour UAE · Indian

91

2.5 yrs

–

Shortlist

3

Maria Santos Cruz
Spinneys · Filipino

88

2 yrs

–

Shortlist

4

Fatima Al-Qasimi
Landmark Group · Egyptian

82

1.5 yrs

✓

Shortlist

5

Ravi Kumar Sharma
Noon.com · Indian

79

1 yr

–

Shortlist

6

Amina Yusuf Hassan
Zoom Supermarket · Somali

76

1 yr

✓

Shortlist

7

Chen Wei Ling
Emax Electronics · Chinese

74

1.5 yrs

–

Review

8

Kofi Asante Mensah
Union Coop · Ghanaian

71

0.5 yr

–

Review

9

Nadia Petrov Volkov
IKEA Dubai · Russian

68

1 yr

–

Review

10

Lakshmi Patel Reddy
Choithrams · Indian

65

1 yr

–

Review

Section 05 · Behind the build

How it was done — step by step.

The DIY Recruitment workflow was built the same way the DIY Marketing coach was — by stacking Claude Code with a few open libraries and a clear set of prompts. Here's the build log, in order.

Step 01 · 15 min

Set up the working folder.

Create the three folders the pipeline writes into. Indeed CVs and WhatsApp CVs land in separate sub-folders so the agent can tag the source in the final JSON.

~/Downloads/indeed-cvs/
~/Downloads/whatsapp-cvs/
~/Downloads/all-cvs/ — merged pool

Step 02 · 10 min

Install Claude Code + the four Python helpers.

Claude Code orchestrates the agent. Python handles the file plumbing — PDF text, DOCX, images and ZIP extraction.

npm i -g @anthropic-ai/claude-code
pip install pymupdf pdfplumber python-docx pillow pytesseract playwright openpyxl
playwright install chromium
Drop your ANTHROPIC_API_KEY into .env

Step 03 · 45 min

Run the Indeed browser agent.

Paste the browser-agent prompt into Claude Code and point it at your Indeed employer dashboard. It logs in, walks the applications list page by page, clicks Download CV on each row, and saves the file with the candidate's name.

Headless Chromium via Playwright
~1,000 CVs in ≈ 45 minutes
Handles pagination + rate-limit back-off

Step 04 · 5 min

Unzip the WhatsApp archive & merge.

WhatsApp's "Export chat" produces a ZIP of every attached file. Unzip it, then copy everything (PDFs, DOCX, images) into the merged pool.

Stage 1 · Ingest WhatsApp $ unzip ~/Downloads/whatsapp-cvs.zip -d ~/Downloads/whatsapp-cvs/ $ cp ~/Downloads/whatsapp-cvs/*.pdf ~/Downloads/all-cvs/ $ cp ~/Downloads/whatsapp-cvs/*.docx ~/Downloads/all-cvs/ $ cp ~/Downloads/indeed-cvs/*.pdf ~/Downloads/all-cvs/ $ echo "Total: $(ls ~/Downloads/all-cvs/ | wc -l)"

Step 05 · 5 min

Paste the JD & lock the rubric.

Drop your Indeed.ae Promoter JD into the agent. Claude extracts requireds, preferreds and deal-breakers, and proposes the 7-criterion weighted rubric (see Section 03). Tweak the weights, then approve.

JD → rubric translation runs in < 10 seconds
Weights can be edited inline before scoring starts
Saved to rubric.json for the audit trail

Step 06 · ~45 min

Run extraction across all 1,040 CVs.

One Claude Code command kicks off the full extraction loop. PDFs are parsed for text; scanned PDFs and image CVs are OCR'd; the agent writes 15 fields per candidate.

Stage 2+3 · Claude Code command $ claude "Process all CVs in ~/Downloads/all-cvs/. Extract: name, email, phone, nationality, UAE experience years, retail background (yes/no), languages, last employer, visa status. Job description: [PASTE YOUR JD HERE]. Score each 0–100 against the promoter rubric. Output a ranked candidates.json file."

Step 07 · 10 min

Build the searchable HTML dashboard.

Second Claude Code command reads candidates.json and writes a self-contained dashboard. Search, eight filters, sortable table, top-10 highlight, Excel-export button.

Stage 4 · Build dashboard $ claude "Read candidates.json. Build a self-contained HTML file with: - Search bar (name, nationality, employer) - Filter buttons: Top 10, Shortlist 75+, Review 60-74, Arabic Speaker, Visit Visa - Sortable table: rank, name, nationality, UAE exp, languages, visa, score, status - Top 10 highlighted in green - Excel export button - Click row to see full CV summary Save as cv-screening-results.html"

Step 08 · 20 min

Review the top 10 & ship the interview invites.

Open the HTML in any browser. Sanity-check the top 10 against the original CVs (one-click view). Export to Excel for HR records. Send WhatsApp interview invites from your own number — never from the agent.

2 minutes per top-10 candidate sanity check
Excel export for compliance / archive
You write the interview invite — human in the loop

Why this works. The same pattern that powers azizsaif.com/diy-marketing — a thin orchestrator (Claude Code), a fat schema (JSON), and a single-file deliverable (HTML) — is the spine of every DIY workflow on this site. Recruitment is just the next vertical.

Section 06 · Technology

The stack — six pieces, mostly free.

Everything is either bundled with Claude Code, free on your existing accounts, or open-source. Total recurring cost: one Claude Pro seat.

CC

Claude Code

Browser agent · orchestrator

Drives Playwright to download CVs from Indeed.ae. Orchestrates the full agent loop. Writes the output HTML. Your CLI AI worker.

AI

Claude API

Extraction · scoring

Reads each CV (PDF / DOCX / image), extracts 15 structured fields, scores against the rubric. Runs in batch — ~1,000 calls via the Anthropic API.

PY

Python

File processing · glue

PyMuPDF / pdfplumber for PDFs. python-docx for Word. Pillow + pytesseract for image CVs. zipfile for the WhatsApp archive.

PW

Playwright

Browser automation

Headless Chromium for Indeed.ae bulk download. Logs in, navigates the applications list, downloads every CV without a single click from you.

XL

openpyxl

Excel export

Writes a formatted XLSX workbook from the ranked candidate JSON. Colour-coded score bands. One-click export from the HTML dashboard.

HT

HTML + JS

Dashboard

Self-contained single HTML file. No server. Open in Chrome / Edge. Search, filter, sort 1,040 candidates. Print-ready. Share by WhatsApp.

Cost estimate. 1,040 CVs × ~500 input + 200 output tokens ≈ 730K tokens total. At Claude Haiku pricing, total API cost lands at $0.20–$0.50 USD for the full batch — effectively free.

Section 07 · Execution timeline

From zero to interviews booked — one afternoon.

Five lanes, one day. Setup, ingest, extract, score, ship. By Day 2 you're on WhatsApp with the top 10.

Day 1 Setup

Install Claude Code CLI
Install Python + libraries
Create folder structure
Test API connection

Day 1 Ingest

Run Indeed browser agent
Download ~1,000 CVs
Unzip WhatsApp archive
Merge into /all-cvs/

Day 1 Extract

Paste job description
Run extraction agent
AI reads all CVs (~45 min)
Generate candidates.json

Day 1 Score

Review scoring rubric
Adjust weights if needed
Run scoring agent
Ranked list generated

Day 1 Ship

HTML dashboard built
Review top 10
Export Excel if needed
Send interview invites

Total time. Setup 1 hr · Indeed download 45 min · AI extraction 45 min · Scoring + output 15 min. Everything done in one afternoon. Ten interviews scheduled by Day 2.

Want this run for your next role?

Bring me the JD. I'll send you a ranked shortlist.

DIY Recruitment runs in your account, on your data, under your control. If you'd rather watch over my shoulder while it runs, send the JD and your application source over WhatsApp — same playbook, run for you.

WhatsApp the JD → See DIY Marketing →