كيف بنينا HR Smart من الصفر باستخدام الذكاء الاصطناعي؟

Evidence A — owner-supplied first Lovable prompt

حدد في البرومبت: الهدف، التقنيات، الوحدات، الأمان، اللغة، القيود الأخلاقية، قيود الخطة المجانية، وأول مرحلة. جملة «Build and test one module at a time» هي محور المنهجية كلها.

Create HR Smart, a personal AI-assisted recruitment management web application.

Use React, TypeScript, Tailwind CSS and Supabase Free.

Build a professional responsive dashboard with Jobs, Candidates, CV Upload, Analysis Results and Settings.

Implement secure authentication, job CRUD, candidate management, private CV storage and user-scoped database policies.

Prepare a modular matching engine with keyword matching and an optional Gemini API integration through secure backend functions.

Support Arabic and English, including RTL layout.

Never automatically reject applicants. Show evidence and uncertainty for every matching result.

Keep the application compatible with free-tier limits. Build and test one module at a time. Start with authentication, database tables, Dashboard and Jobs management. Use fictional data only.

المطلوب:

مرحلة أولى: المصادقة، جداول قاعدة البيانات، Dashboard وإدارة الوظائف — ببيانات وهمية فقط.

النتيجة المسجّلة:

أعلن Lovable تنفيذ تسجيل الدخول وقاعدة البيانات وDashboard والوظائف وSettings مع العربية/الإنجليزية وRTL، وجداول jobs وcandidates وcv_files وanalysis_results بسياسات خصوصية لكل مستخدم. نجحت اختبارات قواعد نموذج الوظيفة والترجمات، لكنه صرّح بأنه لم يختبر الصفحات المحمية يدويًا لعدم وجود حساب.

Evidence B — Claude transcript lines 1–8

افحص قبل أن تعدّل، واطلب خطة قبل الكود، واطلب من الأداة أن تقول بوضوح ما لا تستطيع الوصول إليه.

I am continuing development of HR Smart, originally built with Lovable.
My GitHub repository is hatemkhatib96-web/talent-compass.
Please inspect the project files currently available to you.
First, identify which files you can actually access and which important files are missing.
Then analyze the existing architecture, authentication, database integration, and Jobs module.
My next goal is to implement Candidates Management and CV Upload while preserving the existing Arabic/English design and Lovable Cloud backend.
Do not rebuild the application from scratch, change the database blindly, or use paid services.
For now, provide an implementation plan only. Do not modify any code yet.

المطلوب:

خطة تنفيذ فقط بعد فحص فعلي للملفات المتاحة — دون تعديل أي كود.

النتيجة المسجّلة:

تدقيق فقط؛ راجع Claude في البداية نحو 115 ملفًا مرفقًا ولم يدّعِ وصولًا مباشرًا لقاعدة البيانات.

Evidence B — Claude transcript lines 94–113

لاحظ القرارات المسبقة (تحذير التكرار دون منع، لا OCR الآن) والسؤال الصريح: هل يمكنك فعلًا الـcommit؟ و«لا تدّعِ النشر ما لم يحدث».

Go ahead with Step 0 and Step 1.
My decisions:

CV Upload should have its own sidebar page, and uploading should also be available from each candidate's profile. Later, I want to support bulk uploads of up to 20 CVs.
No OCR for scanned PDFs for now. Show a warning when text cannot be extracted.
Warn about duplicate candidates with the same email, but do not automatically block them.
Important requirements:

Preserve the existing HR Smart application and its current design.
Keep Arabic and English support, including RTL.
Reuse the existing Lovable Cloud database.
Do not create a new database or Supabase project.
Do not modify generated migration files.
Keep the implementation compatible with free online services.
Never automatically reject candidates based on matching results.
Start with Step 0 verification and Step 1 Candidates Management only.
Before writing code, explain whether you can actually modify and commit files to my GitHub repository.
If you cannot, provide complete, ready-to-apply code changes with exact file paths.
Do not claim that changes have been deployed unless they actually have been deployed.
After completing Step 1, show me how to test Candidates Management before proceeding to CV Upload.

المطلوب:

Step 0 للتحقق ثم Step 1 لإدارة المرشحين فقط، مع خطوات اختبار قبل CV Upload.

النتيجة المسجّلة:

عُمل على فرع feature/candidates-management؛ دُفع التنفيذ الأول دون تشغيل مجموعة الاختبارات الكاملة.

Evidence B — Claude transcript lines 187–197

«لا تدمج ولا تنشر» + «إن فشل أي فحص فأصلحه على الفرع» = حلقة Fail → Diagnose → Fix → Re-run.

I successfully created Pull Request #1 for the Candidates Management module.
Before merging, please set up free GitHub Actions CI on the existing feature/candidates-management branch.
Requirements:

Inspect package.json and the lockfile to identify the correct package manager.
Create a GitHub Actions workflow that runs tests, TypeScript checks, lint, and production build.
Make sure generated routes are handled correctly.
Commit and push the workflow to the same feature branch.
Do not merge into main or deploy anything.
If any check fails, investigate and fix the code on the feature branch.
Please provide the GitHub Actions run link and tell me whether all checks passed.

المطلوب:

Workflow يشغّل الاختبارات وTypeScript وlint وbuild على نفس الفرع، مع رابط التشغيل.

النتيجة المسجّلة:

انتهى CI بنجاح 19 اختبارًا وbuild وtypecheck بعد إصلاح اقتباس YAML ودين التنسيق واختبار الأحرف العربية.

Evidence B — Claude transcript lines 246–253

عند نفاد الرصيد: قسّم العمل بين ما يمكن فعله يدويًا وما تكتبه الأداة، واطلب تحقيقًا قبل التغيير.

Important update:
Lovable currently has 0 available credits.
Phase 2 Candidates Management has been successfully tested, and adding candidates works correctly.
Our goal is to continue Phase 3 with minimal cost.
Prioritize Claude + GitHub for development.
First, investigate whether the existing Lovable Cloud Storage and database security configuration can be completed manually without using Lovable AI credits.
Do not make any changes yet.
Provide the safest free implementation path, including any limitations or steps that require Lovable Cloud access.

المطلوب:

أسلم مسار مجاني لإعداد التخزين والأمان، مع القيود والخطوات التي تتطلب وصولًا لـ Lovable Cloud.

النتيجة المسجّلة:

رصيد Lovable كان صفرًا؛ استكشف Claude الإعداد المجاني دون تعديل الكود.

Evidence B — Claude transcript lines 259–263

برومبت قصير يحافظ على القيود نفسها (لا تعديل، لا دفع، لا نشر) ويطلب الإيجاز لتقليل التكلفة.

Continue the previous read-only investigation.
One command failed in the previous attempt. Identify the reason and retry safely if needed.
Check the existing Lovable Cloud Storage setup and determine whether Phase 3 can be implemented without Lovable AI credits.
Do not modify code, push commits, or deploy anything yet.
Keep the response concise to minimize token usage.

المطلوب:

تحديد سبب فشل الأمر السابق وإعادة المحاولة بأمان، بإيجاز لتوفير الـtokens.

النتيجة المسجّلة:

سبب الفشل غياب مجلد supabase/migrations — فشل ls غير ضار.

Evidence B — Claude transcript lines 298–317

طلب تحقيق أمني قبل التنفيذ، وفصل صريح بين «صمّم الاختبار» و«شغّله».

We have successfully configured Lovable Cloud Storage manually without using Lovable AI credits.
Confirmed:

Private bucket: cv-files
Maximum file size: 10 MB
Allowed MIME types: PDF, DOCX, TXT
Four storage.objects RLS policies exist for authenticated users: SELECT, INSERT, UPDATE, DELETE.
Storage policies restrict access to paths beginning with the authenticated user's UUID.
Next task: Security verification before Phase 3 implementation.

Review the existing cv_files table RLS policies and candidate ownership checks.
Identify any cross-user access vulnerabilities.
Provide read-only SQL queries to verify the live database configuration.
Design a safe two-user isolation test for Storage and database records.
Check whether additional policy hardening is required.
Do not modify migrations or generated files.
Do not commit, push, merge, or deploy anything yet.
Our priority is security, minimal cost, and continuing development through Claude + GitHub without Lovable AI credits.
After verification, prepare the implementation plan for Phase 3: CV Upload, PDF/DOCX/TXT text extraction, Arabic/English support, and CI tests.
Do not start implementation until I approve.

المطلوب:

مراجعة RLS لـ cv_files وملكية المرشح، استعلامات SQL للقراءة فقط، وتصميم اختبار عزل بحسابين، ثم خطة المرحلة 3.

النتيجة المسجّلة:

bucket خاص cv-files جُهّز يدويًا؛ نبّهت المراجعة إلى مخاوف تحقق ملكية السجلات الأم. هذه كانت خطة — ليست دليلًا على تشغيل اختبارات العزل.

Evidence B — Claude transcript lines 420–460

لاحظ تحديد مسار التخزين {user_id}/{candidate_id}/{uuid}.{ext} وقيود «لا تُضعف RLS» و«لا ميزات غير ضرورية».

I approve starting Phase 3 — CV Upload & Management.
This project is an educational AI development demo, not a production application. We will postpone additional database security hardening and two-user isolation testing for now.
Current setup:

Lovable Cloud Storage bucket cv-files is configured.
The bucket is private.
Maximum file size: 10 MB.
Allowed formats: PDF, DOCX, TXT.
Four authenticated Storage RLS policies are configured.
Candidates Management is already working.
Your task: Implement Phase 3.

Create a new branch: feature/cv-upload.
Implement CV upload for existing candidates.
Support PDF, DOCX and TXT files.
Extract CV text using pdfjs-dist and mammoth.
Support Arabic and English.
Add CV upload, list, download and delete functionality to candidate profiles.
Create a functional CV Upload page in the sidebar.
Show upload progress, success and error messages.
Allow users to view and edit extracted CV text.
Handle unreadable or scanned PDFs gracefully without OCR.
Use the existing database schema and private Storage bucket.
Follow the existing UI design and responsive layout.
Add appropriate tests and run CI.
Important constraints:

Do not modify database migrations.
Do not weaken or bypass existing RLS policies.
Use the storage path format {user_id}/{candidate_id}/{uuid}.{ext}.
Use fictional CVs for testing.
Keep implementation simple and suitable for a demo.
Do not add unnecessary features or paid services.
GitHub workflow:

Implement on the feature branch.
Commit and push your changes.
Open a Pull Request to main.
Do not merge the PR or deploy automatically.
Report the CI results and any remaining limitations.
Start implementing Phase 3 now.

المطلوب:

فرع feature/cv-upload: رفع/عرض/تنزيل/حذف، استخراج نص بـ pdfjs-dist وmammoth، صفحة جانبية، اختبارات وPR دون دمج.

النتيجة المسجّلة:

نُفّذ PR #2 ونجح CI النهائي. لا OCR ولا رفع جماعي لـ20 ملفًا في هذه المرحلة.

Evidence B — Claude transcript lines 554–650

برومبت تنفيذ مباشر بلا تخطيط بعد أن ثبتت البنية — مع قيود العدالة: لا اسم ولا جنس ولا عمر ولا جنسية ولا صور في التسجيل، والنتيجة استشارية.

Phase 4 — Implement Job Matching Engine
Phase 3 is successfully merged, deployed, and manually tested on Lovable. CV Upload, text extraction, download, editing, and deletion are working.
Start implementing Phase 4 immediately. Do not perform a preliminary repository audit, security review, or planning phase. Do not ask questions. Keep token usage minimal.
Goal
Build a fully functional, free, rule-based CV-to-Job Matching Engine for HR Smart.
Use the existing project architecture, database tables, UI components, and bilingual support.
1. Matching Engine
Implement a deterministic matching algorithm that compares:

Extracted CV text
Candidate skills
Job title
Job description
Required job skills
Candidate experience
Return:

Overall Match Score (0–100%)
Matched Skills
Missing Skills
Experience Match
Keyword Match
Clear explanation of the score
Use a transparent weighted scoring system:

Skills Match: 50%
Experience Match: 20%
Keyword Relevance: 20%
Job Title Relevance: 10%
Normalize scores fairly when some job requirements are unspecified. Do not invent missing qualifications.
Support Arabic and English text, including basic normalization and common skill synonyms.
2. Database Integration
Use the existing analysis_results, candidates, jobs, and cv_files tables.

Save matching results to the database.
Associate each result with its candidate and job.
Allow users to rerun matching.
Prevent duplicate results when possible.
Respect existing RLS policies.
Do not modify migrations or database schema.

Matching UI
Activate the existing Analysis Results page.
Display:

Candidate name
Job title
Match percentage
Skills matched and missing
Score breakdown
Explanation of results
Analysis date
Provide an Analyze Candidate action that lets the user select a candidate and job.
Add clear empty, loading, success, and error states.
4. Candidate Profile Integration
Add a Match with Job button to candidate profiles.
Allow selecting an existing job and viewing the resulting analysis.
5. UI & Localization

Follow the existing HR Smart design system.
Fully responsive.
Arabic RTL and English LTR.
Reuse existing components and translation patterns.
No unnecessary redesign.

Important Constraints

No paid APIs.
No OpenAI API key or external LLM required.
No backend AI services.
No unnecessary dependencies.
No database migrations.
Use fictional test data only.
Results are advisory, not automatic hiring decisions.
Do not use names, gender, age, nationality, photos, or other protected/personal attributes for scoring.
Avoid presenting the score as an objective measure of candidate quality.

GitHub Workflow

Start from the latest main.
Create branch feature/job-matching.
Implement the complete feature.
Add focused unit tests for matching calculations, Arabic/English normalization, missing data, and edge cases.
Run existing CI checks.
Fix errors.
Commit and push.
Open Pull Request #3 targeting main.
Do not merge or deploy automatically.
Final Response
Keep your final report short. Include:

PR link
CI status
Features implemented
Any limitations
Manual testing steps
Begin coding now. No preliminary analysis, no approval checkpoint, and no unnecessary explanations.

المطلوب:

محرك حتمي مرجّح 50/20/20/10 يحفظ النتائج في analysis_results، مع اختبارات وPR #3 دون دمج.

النتيجة المسجّلة:

نُفّذ PR #3؛ 68 اختبارًا إجمالًا منها 30 للمطابقة؛ محرك حتمي مرجّح بلا API مدفوع.