Applyist: An AI-Powered Job Application Assistant

The Problem

Tailoring a resume for every job application is tedious but necessary in today's world. Most people either send the same generic resume everywhere (lower response rates) or spend 30-45+ minutes customizing each one. Cover letters are even worse — high effort, low differentiation.

I wanted to build a tool that handles the repetitive customization work while keeping the human in control of the final output.

What I Built

Applyist is a full-stack web application built with React, Supabase, and the Anthropic Claude API. The core loop:

1. Upload your base resume (PDF upload with text extraction)
2. Add a job application (paste the description or parse it from a URL)
3. Generate a tailored resume — Claude analyzes the job description and rewrites relevant bullets while preserving your authentic voice
4. Generate a cover letter — uses the tailored resume + job context to produce a conversational, role-specific letter
5. Edit inline and download as PDF — every generated output is editable before export

Technical decisions that mattered

Claude API with tool_use for structured output: Rather than asking the LLM to return raw text and parsing it, I used Claude's tool_use feature to get structured JSON responses (experience entries, skills categories, tailored bullet indices). This made the output predictable and directly renderable.

Prompt engineering for tone: The default AI writing voice is stiff and buzzword-heavy. I iterated extensively on system prompts to produce natural, first-person writing. Specific bans: no em dashes, no "spearheaded," no "passionate," no "results-driven." The prompt tells Claude to keep the candidate's authentic voice.

Edge Functions for AI calls: All AI processing runs in Supabase Edge Functions (Deno), keeping API keys server-side and enabling proper error handling for rate limits and billing issues.

Job URL parsing: Built a pipeline using Jina AI Reader to fetch job page content, then Claude to extract the job description, company, and role title. Handles most job boards; graceful fallback to manual paste when it can't.

The Outcome

Applyist is live and functional — I use it for my own job applications. The project demonstrates hands-on AI product development: not just prompting a model, but building the full system around it (auth, storage, structured output, inline editing, PDF generation, error handling).

What I Learned

Building with LLMs taught me that the hard part isn't the API call, it's everything around it. Prompt engineering is product design. Structured output is UX. And the "last mile" (inline editing, PDF formatting, error states) takes 80% of the development time.