Job Hunter Agent

Job Hunter Agent is an open-source agent that finds job postings, writes a tailored résumé for each one and fills in the application form. Nothing is sent until the user approves it.

Role
Independent developer
Timeline
October 2026
Built with
Claude API, Ollama or vLLM, Jev
Links
Source on GitHub ↗
01 · Background

People with several skill sets have to rewrite their résumé and refill a long form for every job they apply to.

Many people now work across several fields, such as design, engineering and strategy, and each job asks for a different part of that experience. In the United States, applicants send dozens of applications, each with a long form that repeats the résumé, and a better response rate means adjusting the résumé for every posting. Most of that work is the same each time, and AI writing tools that shorten it write in a generic voice.

02 · The storybook

One storybook, one tailored résumé per posting, one approval.

One storybook

Users put all of their experience and skills in one file, the storybook. For each posting, the agent reads the storybook against the job description and writes a tailored one-page résumé, using only what the storybook contains.

The form, filled

The agent fills in the application form and takes a screenshot of it, without submitting.

One approval

The résumé and the form screenshot go to a review queue. Nothing is sent until the user approves.

The user’s own voice

The agent writes in the user’s voice, learned from their own writing, and blocks typical AI phrases such as “leverage” and “seamless”.

The storybook with the Experience file open: a role, its context and two résumé lines, each ending with the lenses it fits
The storybook with one file open: each résumé line ends with the lenses it fits. Shown with demo data.
A one-page résumé for Alex Rivera: experience at Fieldwork, a selected project, education and skills
The one-page résumé the agent wrote for Northwind Energy, using only what the storybook contains.
The Northwind Energy application form, filled in: name, email, portfolio, why Northwind, relocation yes, visa sponsorship no, full-time and contract ticked
The application form as the agent filled it, in the screenshot it takes without submitting.
03 · Review

Drafts are ranked by match, and questions only the user can answer come first.

Postings under a 50% match never reach the queue. Each draft shows the résumé, the filled form, salary and location. Approval stays locked until the user has answered every open question, and every step is logged.

Four drafts ranked by match: Northwind Energy 94, Lumen Labs 86, Studio Kite 82 and Fernhill 68 with one question
Drafts ranked by match: green above 90, blue above 80, orange above 50.
The Northwind Energy draft: location, salary, links, the résumé and the filled form, the fields the agent filled, and Approve and Reject
One draft: salary and location, the résumé and the filled form, the answers the agent gave, and the approval.
A question for the user, Are you available for a paid test project, above the approval bar, which is locked
A question only the user can answer comes first, and approval stays locked until it is answered.
An email drafted to Studio Kite, with its recipient, subject and message
For a posting that takes applications by email, the agent drafts the email instead.
The activity log: an application sent, 38 postings rated, 142 postings scraped, and résumés composed for several lenses
The activity log of every step, from finding postings to sending an application.
04 · Learning from notes

Each rejection or note becomes one proposed fix.

A fix can be a better example in the writing brief, a reworded storybook line or a phrase the writing check should catch. The user applies or dismisses it, and the agent’s prompts never change without approval.

The agent’s architecture: the user’s steps in green, the agent’s in white, outside tools dashed
The architecture. The user’s steps are green, the agent’s are white and outside tools are dashed. Opens at full size.
05 · Models

Each step is one structured model call, so the model can be swapped.

The agent runs on Claude, on open-source models through Ollama or vLLM, or on a Claude Code login. Jev from TypeSafe ranks postings. Checkable rules, such as page length and working links, are enforced in code, not in the prompt.

Next projectCodeCut