Markdown prompt files for job search are a sound pattern with unverified claims

The premise here is simple: markdown files that encode domain knowledge and workflow heuristics can be dropped into an AI coding assistant's context, turning a general-purpose tool into something that behaves like a specialist. ResumeSkills applies that pattern to job searching, packaging twenty discrete skills covering everything from ATS keyword matching to salary counter-offer scripting.

The mechanism is worth understanding clearly. Each skill is a markdown file. When you install them - either globally via the npx skills add command or by manually copying files into your agent's skills directory - the AI assistant can pull the relevant file into context when it detects you're working on a matching task. There's no plugin architecture, no API, no compiled code. The "intelligence" is entirely in the instructions written into those markdown files, and the agent's ability to recognize which one applies.

That's both the appeal and the honest limitation. The README makes claims that deserve scrutiny: 2-3x more interviews per application, two months saved on average job search time, salary offers running ten thousand dollars higher. None of these figures are sourced. They read as marketing copy, not measured outcomes. Someone building a serious workflow around this should treat those numbers as aspirational framing, not benchmarks.

What the repo does offer concretely is coverage. The twenty skills span a sensible taxonomy: ATS formatting, bullet rewriting, job description gap analysis, STAR story generation for interviews, offer comparison, LinkedIn sync, and specialized variants for technical roles, executive positions, academic CVs, and career changers. The resume-quantifier skill - which explicitly handles the case where a user doesn't know their actual metrics and needs to estimate - is a pragmatic acknowledgment of how resume writing actually works in practice.

The installation path is clean. A single command installs everything globally across projects, and the skills work with Cursor, Claude Code, Windsurf, Codex, Gemini CLI, and over thirty other agents according to the README. That breadth matters: the pattern isn't locked to one tool.

The deeper question this repo raises is about the skills-as-markdown pattern itself. When the "skill" is just a well-structured prompt with domain context, the quality ceiling is entirely determined by how well the markdown was written - and that's invisible from the outside. You can't audit a markdown file the way you'd audit a library. The value is real but opaque, and the repo gives no way to evaluate individual skill quality before deploying it.