Most AI-powered job search tools are limited to resume creation. Users still have to copy data into application forms by hand, which takes hours. The open-source AI Job Search framework on GitHub automates the entire process locally. It finds relevant job openings that match the candidate's profile and optimizes the resume.

Purpose and Operating Principle

The framework turns a language model into a personal job-search agent that runs on the user's computer. It performs three tasks:

• finds current job openings on connected job portals;
• removes unreliable and duplicate listings;
• prepares verified job applications.

The author built the tool for their own job search and became fully employed using this process.

According to the author, when tested on 69 tailored applications, the agent received 21 invitations to a first interview. The conversion rate was 29%.

Prompt Injection Protection

Public job descriptions may contain hidden instructions designed to confuse automated parsers. Version 1.7 treats all external descriptions as raw data. Before passing information to an AI agent, the framework filters out non-standard formatting.

Resume Evaluation

The framework evaluates resumes on an objective scale from 0 to 100. Scores are reduced if the resume lacks the required skills listed in the job posting.

Requirements and Setup

• Hardware. The framework runs on standard laptops with 4 GB of RAM. Cloud servers are not required.
• Privacy. Contact data and salary information stay on the local machine and are not sent to external cloud services.
• Runtime environments. Claude Code is not completely free. Alternatives include Codex, Antigravity, and Gemini CLI.
• Instructions. All commands and instructions are provided in the agents.md file in the official repository.

The author used Google Antigravity. The service is free and has daily limits on agent usage.

Model Selection

Antigravity offers the Gemini 3.8 Flash model. It has three thinking modes: low, medium, and high. The medium and high modes produce more accurate results but take more time. The author chose the low mode to speed up testing.

Installation Process

The author gave the agent a prompt to study the repository, check the instructions in agents.md, and execute the official commands. After installation, the agent initialized the skills file and verified all the executable commands that the skill suggests.

Workflow

Stage 1. Data Collection and Cleaning

The agent receives the target roles and salary expectations, extracts skills from the input files, and queries the connected job portals. It then removes duplicates and dangerous instructions that may be embedded in job postings, and compiles a list of current listings to respond to.

Stage 2. Match Assessment

Job postings are evaluated for technical fit and salary requirements. The final score is expressed as a percentage from 0 to 100. Postings older than 21 days receive penalty points. This makes it possible to avoid wasting time on outdated listings.

Stage 3. Document Preparation

A cycle of two agents operates:

• The drafting agent creates a resume and a cover letter tailored to the specific job posting.
• The review agent checks whether all job requirements are met and removes excess keywords.

Profile Setup

The /setup command loads the base resume and saves the target salary and projects to configuration files in JSON format. The files are stored locally in the agent's folder.

The author used a candidate JSON file with the fictional name Private Ray. The profile contains:

• target roles;
• location preferences;
• core skills;
• experience with the current stack;
• a focus on autonomous multi-agent pipelines.

Framework Commands

Commands can be run individually if you need to specify particular parameters.

Framework Operating Rules

1. Human responsibility. The agent only prepares documents. The user performs the final submission.
2. Data protection. Employment information is stored only in private branches.
3. Search modules. Modules can be replaced to search for vacancies on domestic or international platforms.
4. Quality over quantity. Weak matches are filtered out in advance to increase the share of applications that employers respond to.

Job Search Results

The agent found four job postings filtered by the "remote work" criterion:

1. AI systems engineer, Cognitive Flow Labs. Remote work for the US or Europe. The posting was published 3 days ago on GitHub Rising Jobs. The salary range, required skills, and experience are specified. The agent checked the description for prompt injection instructions and found no hidden commands. Rejected insertions are noted in the report.
2. Second and third postings. One of them relates to an architect role.
3. AI infrastructure engineer. The fourth posting.

Оценка соответствия

Агент сравнил вакансии с резюме автора и рассчитал итоговый балл. Штрафы назначались за отсутствие обязательных требований.

Автор установил порог соответствия 75%. Вакансия AI infrastructure engineer получила 68% и была отклонена. Основная причина: в резюме отсутствовал навык работы с кластерами Kubernetes.

По вакансии Cognitive Flow Labs агент отметил следующие показатели:

• зарплата полностью соответствует требованиям;
• опыт в предметной области превышает требуемый;
• технические навыки полностью совпадают с резюме;
• штрафы не назначены;
• вакансия опубликована недавно.

Preparing the Application

For the Cognitive Flow Labs position, the agent prepared the following materials:

• Adapted resume. The professional summary, skills section, and descriptions of experience and key projects were revised. Skills from the job posting were moved to the beginning of the document to get past ATS systems.
• Cover letter. The letter expresses interest in the role and describes prior experience and an open-source project the author built. The closing includes the name, current job title, email address, and a link to the portfolio.
• Audit report. A separate agent checked the data in the resume against the information provided by the author.

Preparing all materials took less than two minutes.

Comparison with General-Purpose Methods

General-purpose AI models often lose focus when processing random web text and produce incorrect responses. The framework isolates raw text from system instructions and performs operations in a secure sandbox.

General-purpose models can create résumé bullet points that were not in the candidate's experience. The tool links each bullet point to the user's verified experience.

Testing Results

All results are verified by a reviewer agent.

Advantages

• Security isolation. An independent sandbox blocks the execution of dangerous instructions in the user's system.
• Honest assessment. The framework lowers scores for gaps in skills. Earlier models inflated compliance metrics.
• Two-agent verification. The system checks for formatting errors and removes skills that the agent added to its experience without confirmation.

Limitations

• Users must review every response before sending it.
• Local job portals require separate parsing plugins.
• The quality of the final documents depends on the completeness and accuracy of the source profile. Career information must be provided precisely and in full from the start.