career-ops: the job search assistant that runs locally in your AI coding CLI, drafting applications only — pressing Submit is always up to you

Who it's for

• Job seekers who want to handle opening evaluation and application tracking inside their own AI coding CLI: hand the assistant a job's description text or URL, and get back an evaluation of that opening, a written report, a saveable PDF file, and an application record in the tracker; the evaluation is there to judge, before you apply, whether the opening is worth your time. If you want to screen openings first and then decide which to pursue, this flow covers exactly that first stretch from evaluation to tracking.
• For people who value open source and making decisions themselves: career-ops is an open-source project, officially positioned as an open-source AI job search assistant, and it states explicitly that every decision is left to the user; it only drafts and analyzes — it doesn't choose for you.
• Who it isn't for, and prerequisites: it runs as a skill inside an AI coding CLI, and whether there's a way to use it independently of such tools is not stated in the evidence; support for non-job-search tasks, specific industries, or particular text languages is equally unknown; the prerequisite for use is a terminal where the official install command can run, and no further requirements for your machine or environment are given officially — confirm those yourself before starting; treat anything left undocumented as unknown, and don't assume capabilities it hasn't demonstrated.

First attempt in 3 minutes

• Prerequisite: a supported AI coding CLI already installed on the machine, and the ability to open a terminal and run commands; you're not tied to any single product.
• Install: run npx @santifer/career-ops init in the terminal. Following the open standard, it writes the skill definition to .agents/skills/career-ops/SKILL.md and creates symlinks or references for each supported CLI (for example .claude/, .qwen/, .antigravitycli/, .grok/). This step writes into those directories — make sure you accept that before running it.
• Trigger: start an analysis task from the CLI conversation: paste a job's description text or URL and ask it to run the official auto-pipeline, whose original wording is “AUTO-PIPELINE: evaluate + report + PDF + tracker (paste text or URL)”.
• Side-effect note: this pipeline generates a PDF and writes tracker entries — that is, it creates local files or records; the exact write location is not specified officially. If you want to see what will be created before it happens, you can ask it to first show the items it plans to create; after the first run, it's also wise to check which new files actually appeared.
• Success signal: the output shows an evaluation and report for that opening, the exported PDF, and a new record in the tracker. All four present means the install-and-trigger chain works and you can move to the next step; if any one is missing, the chain didn't complete — check whether the installation succeeded before retrying.

Use cases

• Case one, evaluating a candidate opening: you have several openings on hand and want to pick the ones worth applying to first. How to start: send the assistant that opening's description text (or URL) and ask it to run the evaluation pipeline above. The real workflow involved is exactly that chain: evaluation, report, PDF, tracker entry. Expected result: you get the opening's evaluation and report, a PDF, and one more record in the tracker; repeat the same steps for other openings, compare several evaluations side by side, then decide which to apply to first. Each new opening adds a record, and the tracker gradually becomes your job search ledger.
• Case two, drafting and filling in an application: you've decided to apply for an opening and want to prepare the materials with less effort. How to start: ask it in the conversation to draft and fill in that application. The capability involved, in the official words, is “It drafts and fills; it never submits. You press Submit.”. Expected result: you get an application draft you can review item by item; it won't submit for you — the last step is always completed by your own hand.

Boundaries

• Side effects and the confirmation point: installation writes the skill definition and creates symlinks or references in each CLI directory (such as .claude/); the evaluation pipeline creates PDFs and tracker entries. Applications are never submitted automatically — officially it only drafts and fills, and the submit action is always performed by the user; this is a confirmation point that must be carried out by a human before applying, not a formality that can be skipped; the point of it: the materials can be prepared by the tool, the decision is always yours.
• Runtime and client prerequisites: officially, it runs on any major AI coding CLI through the open Agent Skill Standard; Claude Code, Codex, Gemini / Antigravity, OpenCode, Grok and Qwen are given as examples with the note that there are more, so it's not a closed roster; all clients share the same skill definition and run locally.
• Explicit official restriction: it never submits applications automatically; submission happens when the user presses. If you expect a tool that finishes applications fully automatically, it doesn't match that expectation.
• Not yet confirmed: the scope of file reads and writes beyond the submission step, whether it runs other commands, whether it accesses the network, exactly where PDFs and tracker entries are written, system dependencies, and uninstallation — the evidence says nothing about these. “Never submits” is only the official guarantee for the application step; it doesn't mean the tool never reads or writes files or performs other operations — judge by the actual behavior on first use.