Open Career Skills

The prompt architecture behind JobMentis, open-sourced as installable Claude skills.

Generic AI career advice gets people rejected: polished emptiness, invented metrics, "results-driven professional" boilerplate that recruiters spot in three seconds. These skills are built the opposite way. Their defining behavior is that they ask you a question rather than invent a fact about you, and they enforce the writing craft (plain verbs, real numbers, say-it-out-loud naturalness) that survives both ATS parsers and human readers.

Run them locally in Claude Code or Claude Desktop, on your own subscription, with your data on your machine.

The five skills

Skill	What it does	The rule that makes it different
ruthless-cv-optimizer	Rewrites your CV against a specific job description	Two-phase flow: it diagnoses and ASKS about suspect numbers and empty bullets, waits for your answers, then rewrites. Gaps are reported honestly, never papered over.
star-story-extractor	Turns a messy braindump into a sharp STAR interview story, filed in your story bank	Missing metric: it asks you ONE question instead of estimating. Also matches your bank against any pasted JD.
linkedin-360brew-planner	Plans and drafts LinkedIn posts under the 2026 ranking rules	Built on LinkedIn's 360Brew decoder model (the reason hashtags and pods died). Saves over likes, analytical first line, pillar discipline.
cover-letter-writer	A 300-400 word letter grounded in your real stories	Two or three deep JD connections instead of six shallow ones; banned-cliche list enforced.
mock-interviewer	A pressure-tested rehearsal, one question at a time	Stays in character, interrupts rambling, drills "what did YOU do?", then debriefs against structure / specificity / ownership / quantification.
job-search	Finds current postings matching your profile	Zero dependencies; only reports jobs it actually fetched or you pasted. Blocked boards get you ready-made search URLs, never invented listings.
evidence-miner	Harvests real achievements from your git history into the story bank	Groundedness rule: no citable artifact (commits, PRs, dates), no proposed achievement. Business impact numbers come from you, never inferred from code.

The /apply pipeline

The skills also compose into one end-to-end flow per job:

The reviewer is a separate subagent that audits every claim in the drafts against your actual profile and stories; anything unsourced gets flagged for removal, not "improved". After the pipeline, /mock-interview rehearses the evaluation's real gaps and /upskill turns gaps recurring across 2+ applications into a learning plan. Full walkthrough: docs/WORKFLOW.md. Per-skill contracts: docs/SKILLS.md. Something misbehaving: docs/TROUBLESHOOTING.md.

Install

Option A: fork this workspace (recommended)

The repo is a ready-to-run Claude Code career workspace with file-based memory: your profile in profile/, your STAR stories in story-bank/, your content calendar in content/.

# fork on GitHub first (keep your fork private if you plan to push), then:
git clone https://github.com/squerne/open-career-skills.git
cd open-career-skills
claude
# then, inside Claude Code:
/setup

/setup builds your profile from a pasted CV or a guided interview. From there:

/find-jobs        # real postings matching your profile, fit-triaged
/apply <url>      # the full pipeline: evaluate, draft, review, track
/optimize-cv      # paste a JD, get a tailored CV in output/
/extract-story    # braindump a win, get a STAR story in story-bank/
/mine-evidence    # harvest achievements from your git history
/plan-content     # 6 post ideas, or draft one post
/cover-letter     # tailored letter in output/
/mock-interview   # one question at a time, debrief at the end
/upskill          # learning plan from gaps recurring across applications

Your personal data (profile/profile.md, your stories, your calendar, everything in output/) is git-ignored by design: even if you push your fork publicly, your CV and metrics stay local.

Option B: install individual skills

Copy any skill folder into your own project or user skills directory:

cp -r .claude/skills/ruthless-cv-optimizer ~/.claude/skills/

Works in Claude Code and Claude Desktop (any agent that reads SKILL.md skills). Without the workspace, skills ask you to paste your CV instead of reading profile/.

Option C: plain chat (ChatGPT, claude.ai, Gemini)

The five writing skills have self-contained paste-able versions in prompts/. Paste one as your first message and follow along. You lose the file-based memory; the prompts tell you what to paste each time. (/apply, /find-jobs, /upskill, and /mine-evidence need file and web tools, so they are Claude Code only.)

Why we open-sourced this

Prompts are a commodity; anyone can copy text. What they can't copy from a markdown file is the system around it. We build JobMentis, a career OS where these same engines run against a persistent, structured version of your career: CV analyses, a searchable story bank, a job tracker, and interview intelligence that feed each other automatically.

This repo is the honest free tier. It genuinely works, and we want it to: a job search run on grounded, non-fabricated assets is better for everyone, including the recruiters. What the hosted product adds is scale and statefulness:

	This repo	JobMentis
The prompt craft	✅ identical philosophy	✅
Your cost	your Claude subscription	free tier + Pro
Memory	markdown files in your fork	structured, searchable, cross-device
CV ↔ story bank ↔ JD matching	manual, per session	automatic, across your whole pipeline
Job pipeline tracking	a folder of files	tracker with per-vendor ATS rules
Mock interviews	text chat	live voice, company-specific question banks
Personality-aware coaching, human coaches	❌	✅

If the flat files start feeling like the bottleneck, that's the moment we built the product for: https://jobmentis.com/en/signup?ref=oss-prompts

FAQ

Do these work with models other than Claude?
The prompts/ versions work with any capable chat model. The SKILL.md versions target Claude Code / Claude Desktop.

Why do the skills refuse to add metrics to my CV?
Because invented numbers are the fastest way to fail an interview. The skills ask you for the real figure; if you don't have one, an unquantified true bullet beats a quantified fake one.

What is 360Brew?
LinkedIn's decoder-only ranking foundation model (published on arXiv, January 2025). It reads post text directly, which is why hashtags, pods, and link-in-first-comment tricks stopped working. Longer explanation: https://jobmentis.com/en/guide/linkedin-algorithm-guide

Is my data sent anywhere?
Only to the model provider you already use (Anthropic, OpenAI, ...). The workspace stores everything in local files, and the .gitignore keeps personal files out of your commits.

Can I contribute a skill?
Yes. See CONTRIBUTING.md. The quality bar: no skill may fabricate user facts, ever.

Acknowledgments / Credits

This repo stands on ideas from two MIT-licensed projects, reimplemented (not copied) for this workspace:

MadsLorentzen/ai-job-search: the fork-and-run workspace pattern, and the concepts behind our /apply drafter-reviewer pipeline, the application tracker, and /upskill gap analysis. If you want LaTeX/PDF CV output or Danish job-board CLIs, use his repo; it does both better than we ever will in Markdown.
Play-New/apply-new: the principle that career evidence should come from work artifacts and that every prose claim should trace to underlying data. Our evidence-miner skill and the reviewer's groundedness audit are descendants of that idea.

License

MIT. "JobMentis" is a trademark of JobMentis; the license covers the prompts and code in this repo, not the brand.