Career-Ops: The Repo That Turns Claude Code Into a Job-Search Operating System

A Job Search, but the Agent Is in Charge

Career-Ops is not trying to be a nicer job board or a faster autofill tool. It treats the repository itself as the interface, with Claude Code reading instructions from files, choosing modes, and moving through a workflow that looks more like an operating system than a script collection.

That inversion is the point. Instead of a person driving software to churn out applications, the repo gives the model a workspace, a policy layer, and a memory layer, then asks the human to review the output before anything goes out the door.

How the Skill Router Decides What Happens Next

The first interesting file is .claude/skills/career-ops/SKILL.md. It acts like a router, detecting whether a pasted URL, a job description, or a request for a specific task should trigger evaluation, PDF generation, or a different mode entirely.

That matters because the system is not one long prompt pretending to be software. It is a routing layer for agent intent, with shared context loaded first and narrower instructions layered on top when the task demands it.

# Routing idea, simplified

- If input is a URL or job post text, enter auto-pipeline
- If the task is PDF generation, load pdf mode
- If the task is application writing, load apply mode
- Always prepend shared context before task-specific instructions
- Keep the active context narrow so the agent stays focused

The design choice is subtle but powerful. Less context is not just cheaper. In this repo, less context is also a way to get better judgment from the model.

The Real Brain Lives in Markdown

Career-Ops pushes the idea of executable docs further than most repos do. The modes/ directory holds task-specific instructions in Markdown, while _shared.md acts as the common base layer that every mode inherits.

That means behavior is distributed across plain text files instead of hidden in a monolith. The agent is not just reading docs. It is effectively reading policy, and that policy is modular.

Why the System Refuses to Spray and Pray

This project is deliberately anti-volume. The scoring logic is built to filter hard, with a North Star metric and weighted dimensions that make the system recommend against weak fits instead of pushing every lead through the pipe.

That restraint is not a disclaimer bolted on for optics. It is part of the product logic, and it shapes what the agent is allowed to generate next.

ATS Optimization as a Systems Problem

The sharp technical detail here is the normalization layer in generate-pdf.mjs. It strips or replaces characters like smart quotes, zero-width spaces, and em-dashes so applicant tracking systems do not choke on polished typography.

function normalizeTextForATS(text) {
  return text
    .replace(/[“”]/g, '"')
    .replace(/[‘’]/g, "'")
    .replace(/—/g, '-')
    .replace(/[\u200B-\u200D\uFEFF]/g, '');
}

That is the story in miniature. The document still needs to look good for people, but it also has to survive a machine path that rewards plain, boring, predictable text.

Markdown as a Database, Tracker as Memory

Career-Ops stores applications in Markdown, reports in Markdown, and configuration in YAML. That keeps the system inspectable, portable, and easy for both the human and the agent to read without a database server in the middle.

Then analyze-patterns.mjs turns those files into a post-mortem layer. It uses text extraction and regex-based parsing to look for correlations between role archetypes and outcomes, which means the repo can learn from its own history without leaving plain files behind.

Applications.md -> Reports/*.md -> Pattern extraction -> Outcome correlations -> Evaluation heuristics

This is where the repo starts to feel like memory, not just storage. The tracker is not a spreadsheet replacement. It is the state of the system.

What Career-Ops Is Really Competing With

Career-Ops does not belong in the same bucket as a single-purpose resume tool. It is closer to a stack that fuses evaluation, generation, tracking, and policy into one agent-native workspace.

That is why the most useful comparison is philosophical, not feature-by-feature. Career-Ops is not trying to win the auto-apply race. It is trying to change the shape of the race.

Why This Repo Matters Beyond Job Hunting

The broader lesson is bigger than recruiting. Career-Ops shows that a repo can become a skill surface for an LLM, that Markdown can serve as state, and that constrained agent systems can be more useful than open-ended chat.

If that pattern spreads, a lot of software will start looking less like apps and more like operating environments for model-driven work. Career-Ops is a strong early example of that shift.