Building Jobisque: An AI-Powered Job Risk Analyzer from Zero to Launch

The Spark: Why I Built a Tool to Analyze Job Listing Red Flags

Every developer has applied for a job that turned out to be nothing like the listing. Unrealistic requirements ("5 years of experience with a 2-year-old framework"), suspiciously vague descriptions, or offers that seem too good to be true.

I wanted to build a tool that could catch these red flags before you waste time applying. That's Jobisque — an AI-powered job risk analyzer.

Tech Stack: Next.js + Claude API + Scoring Engine

Frontend: Next.js (App Router, Server Components)
AI Engine: Claude API (Anthropic)
Scoring: Custom weighted algorithm
Database: PostgreSQL (via Prisma)
Auth: NextAuth.js
Deployment: Vercel

The architecture is simple by design. A user pastes a job URL or description, Claude analyzes it against 15+ risk factors, and returns a detailed risk score with explanations.

Prompt Engineering: Getting Claude to Reliably Score Job Listings

This was the hardest part. Early prompts gave wildly inconsistent results. The breakthrough was structured output:

const systemPrompt = `You are a job listing analyst. Analyze the following
job listing for risk signals. Score each category 1-10.

Categories:
1. Requirements Realism (are they asking for impossible combos?)
2. Compensation Transparency (is pay mentioned? Is it reasonable?)
3. Company Legitimacy (are there verifiable company details?)
4. Role Clarity (is the actual job clear?)
5. Red Flag Language (urgency pressure, vague promises)

Return JSON: { scores: { category: number }, flags: string[], summary: string }`;

Key lessons:

• Structured JSON output makes scoring consistent
• Few-shot examples in the prompt improved accuracy by ~30%
• Temperature 0.3 gives reliable analytical results
• Caching identical listings saves API costs

The Launch: First Users, Feedback, and Iteration

I shipped the MVP in 10 days. The first version was rough — no auth, no saving results, just paste-and-analyze. But it worked.

Initial user feedback shaped the product:

• "Can you analyze LinkedIn job URLs directly?" → Added URL scraping
• "I want to compare multiple listings" → Added a dashboard
• "The scores don't explain enough" → Added detailed breakdowns per category

Metrics

| Metric | Value | |--------|-------| | Time to MVP | 10 days | | Analyses run | 1,000+ | | Average risk score accuracy | ~85% (user-validated) | | API cost per analysis | ~$0.02 |

Lessons Learned Shipping Solo

1. Ship the ugly version first. My MVP had no CSS framework. Users didn't care — they cared about the analysis quality.
2. Claude is remarkably good at structured analysis. The JSON output mode makes it almost like calling a regular API.
3. Prompt engineering IS the product. 80% of the value comes from how you ask the AI, not the UI around it.
4. Solo doesn't mean alone. I used Claude Code as my pair programmer throughout. It caught bugs I would have shipped.

---

Building an AI-powered SaaS? Let's talk. Book a call.