job-search-score

An agent skill for Claude Code that runs a job search the way a careful person would: it verifies before it recommends, checks whether an employer can legally hire you, ranks by how winnable a role is rather than how well it matches, and writes application documents in your own CV format rather than a template it invented.

It is built to reject roles, not to spray applications.

Why this exists

Naive job tooling title-matches a search and ranks by vibes. That produces three specific failures:

1. Wrong company. A company name collides with a better-known one and you assume the wrong sector, size, or business entirely. The fix: read what the job description actually says the company does — never infer identity from a name on a card.

2. Agency listings that look like direct ones. The "company" is a recruitment agency and the description is about an unnamed client. It reads like an employer posting and isn't. Its advertised salary is the client's budget, not a contracted offer.

3. A perfect-fit role you can never win. 200 applicants, or an employer who legally cannot sponsor your visa. Fit alone is not a ranking.

What's actually in here

Three ideas worth stealing even if you don't use this

1. The visa sponsorship gate (UK)

If you need sponsorship, this runs before fit matters. Cross-reference every employer against the UK government's Register of licensed sponsors (a public CSV) and check for the Skilled Worker route specifically.

(Register cross-checking isn't new — see Prior art. The two traps below are the part I haven't seen handled elsewhere.)

The trap: other routes look like a licence and aren't. Global Business Mobility only covers transferring existing overseas staff into the UK — it cannot be used to hire you locally, and it does not lead to settlement. A company can appear on the register and still be a dead end.

The second trap: a B-rated sponsor cannot issue new Certificates of Sponsorship. A licence they can't use is no licence at all — and the register prints the rating right next to the route.

The third trap: a licensed, A-rated employer paying below the visa's salary floor still cannot sponsor you. All of it has to hold at once.

The fourth, which beats all of the above: the job description itself may say "we cannot offer sponsorship for this role." A negative phrase in the posting overrides anything the register says. Check it first — it's free.

And the warning that matters most: never auto-assert a non-exact match. The register lists legal entities; job boards show brands. Every shortcut around this produces confident lies:

sponsor_check.py treats only an exact normalised match as verification. Everything else is proposed with a score and an explicit "this is not verification" warning, for a human to confirm.

A false "they can sponsor you" costs someone years. A missed match costs one email. Optimise for the second error, never the first.

2. Rank on fit × sponsorship × (low) competition

A role can match you perfectly, hold a sponsor licence, and still be hopeless at 170 applicants. Applicant count is exposed in LinkedIn's public response (num-applicants__caption) — most tools ignore it.

The counter-intuitive part: the least contested roles are the ones with friction. An unglamorous domain. A relocation clause. Multi-hour tests. A boring title. That friction is the moat, because most people won't push through it — and a frictionless role (no experience bar, published salary, desirable city) can take 100+ applicants in a day.

3. Clone your CV's format; never design one

build_cv.py opens your existing .docx, keeps its fonts, colours, borders, tab stops and page setup, and swaps in new content. The output matches your CV by construction, not by imitation.

Two things this saves you from:

• textutil -convert docx file.md does not render Markdown. It dumps literal # and ** into the document. If you've ever had a CV come out looking like a draft, that's why.
• Every bullet comes from a verified bullet bank in X-Y-Z form (accomplished X, as measured by Y, by doing Z). Never invent a metric. If it isn't in the bank, it doesn't go in.

4. Catch the AI slop before a human does

voice_check.py <file.docx> reports the mechanical tells that make writing read as AI-generated: em-dash density (the #1 signal — it once hit 14–18 per document), banned slop phrases ("leverage", "passionate about", "seamlessly"…), AI sentence shapes ("not just X but Y"), and punctuation collisions. It reports, it doesn't auto-fix — a real word in context beats a false deletion, so a human makes the call. (Same principle as "report the page count as a fact, don't tune it.")

What a regex can't judge is whether it sounds like you. That's the second pass: copy voice-dna.example.md to voice-dna.md, write down how you actually sound, and check the document against it. The goal is writing a hiring manager reads as your own thinking, not vibe-coded — which is the thing that separates the offers from the rejections.

Setup

Requires Claude Code, Python 3, and macOS or Linux.

git clone https://github.com/aayushiagratha/job-search-score.git
mkdir -p ~/.claude/skills
cp -r job-search-score ~/.claude/skills/

cd ~/.claude/skills/job-search-score
cp config.example.json config.json   # then edit it

Fill in config.json: your CV, your verified bullet bank, your CV template, target titles and locations, salary band, hard disqualifiers, and — if you need one — your visa situation.

Then, in Claude Code:

find me product marketing jobs in London posted this week

It will search, verify, check sponsorship, rank, and stop — presenting everything in chat and building nothing until you pick a role. That pause is deliberate.

Generating documents directly

CV_TEMPLATE=~/Documents/my-cv.docx python3 build_cv.py content.json out.docx

The JSON schema is documented in the script's docstring.

⚠️ Terms of service

This uses LinkedIn's public, unauthenticated jobs-guest endpoints. No login, no cookies, no account credentials — so there is no account to ban. But automated access is against LinkedIn's Terms of Service, and they can rate-limit or block by IP.

Use it at human volumes for your own job search. The pacing and backoff in linkedin-endpoints.md exist for that reason — respect them. This is a considered-search tool, not a scraper, which is also the entire point of it.

Do not point it at an authenticated session. Automating a logged-in LinkedIn account is the thing that actually gets accounts restricted, and it would put at risk the exact account your job search depends on.

Credits

The LinkedIn guest-endpoint approach, the two-pass experience filter, and the pacing/backoff pattern were learned from:

• career-ops (MIT) — the compensation-reliability and posting-legitimacy checks in comp-and-legitimacy.md are adapted from its evaluation prompts.
• ai-job-search — source of the jobs-guest endpoint documentation.
• linkedin-job-scanner (MIT) — the f_E two-pass strategy (untagged postings vanish under a filtered search), the pacing constants, and the reminder that \b word boundaries silently break c++, c#, .net and node.js in keyword matching.

Prior art on the sponsorship gate — read this before assuming it's new

Cross-referencing employers against the UK licensed-sponsor register is not a new idea, and it would be dishonest to imply otherwise. Other people have built it, some of them more rigorously:

• Sponsor-Radar — the closest and most serious. Daily register refresh, trading-name extraction, B-rated sponsors flagged as unable to issue new Certificates of Sponsorship, and a human-in-the-loop override table to kill fuzzy-match false positives. Its README documents the same failure I hit (their trigram matcher paired Encord with "Encortec Limited"; mine paired Tesco with "ATESCO CONSULTANCY LTD"). If you want the register handled properly, read theirs.
• uk_sponsors_inner_join_job_offers, uks, sponsor-list — simpler register lookups and matchers.

What this skill adds on top of that prior art:

• The route trap, named as a trap. Being on the register is not enough. Global Business Mobility only covers transferring existing overseas staff and cannot be used to hire you locally. I have not found another tool that surfaces this distinction as a disqualifier.
• The salary floor, enforced. A licensed employer below the visa threshold still cannot sponsor you. Sponsor-Radar lists this as future work; here it is a gate.
• Competition as a ranking axis. Applicant count is in LinkedIn's public response and, as far as I can tell, no other tool in this space uses it.
• Documents generated by cloning your own CV's formatting, rather than rendering into a template someone else designed.
• It's an agent skill, not a pipeline. No deployment, no database, no hosting. You talk to it.

Licence

MIT.