Jobbot.ai
- AI Labs
- Jobbot.AI
Trouble seeing it? Open JobBot.ai in a new tab
JobBot.AI
An AI job-hunting assistant, built by prompting Claude Code
JobBot.ai turns a resume into a structured profile, scores it against any job posting on a 100-point ATS rubric, and rewrites it for that job without inventing facts. Every evening at 7 pm in each user’s time zone it searches new postings from the Adzuna and USAJobs APIs, ranks them, tailors resumes for the best matches and emails a summary. It’s a full MERN application (MongoDB, Express, React, Node.js) with a background job worker, plan tiers, an admin audit log, and an AWS deployment on ECS Fargate, CloudFront, S3 and SES.
For this project I deliberately changed how I work. Instead of hand-writing code and typing terminal commands, I built and edited the app mainly by prompting Claude Code. I treated it the way a lead treats a developer: I wrote the design, the architecture standards and a 15-milestone plan first, then drove each milestone with short prompts that each had a clear “Done when” check. My job was to set direction, review plans before code was written, test the results and decide what to change. Claude Code even pushed back on parts of my own plan, catching a proxy setting that would have broken behind CloudFront and suggesting safeguards I adopted.
The hardest problem wasn’t getting AI to write a resume. It was stopping it from making up details, and then stopping it from sounding like AI. I designed a tailoring loop where Claude drafts, code fact-checks the draft against the user’s real profile, and a mixed code-and-Claude rubric scores it, with up to three attempts fed by the previous attempt’s feedback. When a recruiter flagged an early draft as AI-written, I prompted a “human voice” layer into place that limits bullets and skills, removes keyword stuffing and penalizes AI-sounding patterns.
The result is roughly 9,500 lines of JavaScript, about 100 automated tests that never call paid services, and a working production app. It showed me that clear specifications and careful review can matter more than typing the code yourself.
Stack: React 19, Vite, Tailwind 4, Node 22, Express 5, MongoDB/Mongoose, Agenda, Claude (Haiku and Sonnet), AWS (ECS Fargate, CloudFront, S3, SES, Secrets Manager)