agentleFS
Miracle-AlignerMiracle-Aligner/job-search-hq/.claude/skills/add-vacancy/SKILL.md

add-vacancy

Skill1 starsChanged 5 months agoMITView on GitHub

Use whenever the user wants to ingest one or more job vacancies into Job Search HQ. Triggers: /add-vacancy, "add this vacancy", "add these jobs", pasting one or more job URLs, or pasting an aggregator/search URL and asking to expand it. Handles a single URL, a bulk list, or an aggregator page (DOU / Djinni / LinkedIn). Loads the candidate profile from `profile/candidate.md`, scores against `docs/rubric.md` via `scripts/score.py`, cross-links the role on other platforms, appends to `data/vacancies.json`, validates, and re-bakes the HTML's embedded JSON block.

The file

We split this skill into the rules it gives an agent

5 rules. About 0 are common, 0 are distinctive, and 5 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈0

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 5

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Read profile/candidate.md first.

    line 80
  2. Very specific:

    Don't eyeball. Don't hand-set tier.

    line 91
  3. Very specific:

    status: default to_apply unless the user states otherwise (e.g.

    line 106
  4. Very specific:

    Never infer a pipeline stage (hr_screen, rejected, etc.) from the posting — only from what the user tells you about their conversations.

    line 109
  5. Very specific:

    Don't duplicate an existing id.

    line 135

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

add-vacancy

Flow

  1. Detect mode — single URL · bulk URLs · aggregator URL.
  2. Fetch each posting (web_fetch for plain pages; Claude in Chrome for JS/auth-gated).
  3. Find alt-source links on other platforms.
  4. Assess against profile/candidate.md; score via scripts/score.py compute.
  5. Preview + confirm with the user.
  6. Append to data/vacancies.json (the v2 envelope's vacancies[]) → python3 scripts/validate.py → python3 scripts/build.py --real.

Modes

InputModeNotes
One job-posting URLSingleStep 1A
Several job-posting URLsBulkStep 1A applied per URL; collect failures, don't abort the batch
A category / search / aggregator URLAggregatorDefer to /refresh-listings (its per-source playbooks live there)

DOU posting URLs look like jobs.dou.ua/companies/<co>/vacancies/<id>/. Djinni: djinni.co/jobs/<id>-<slug>/. LinkedIn: linkedin.com/jobs/view/<id>.

Step 1A — fetch a single URL

  • DOU: plain WebFetch works. Confirm the role title is on the page; if the page redirects to the company's vacancy list or shows "вакансія неактивна", the posting is gone — tell the user and don't add it.
  • Djinni: WebFetch works for individual postings.
  • LinkedIn: WebFetch is auth-gated and usually fails. Use Claude in Chrome with the user's logged-in session.

Extract: company, role, salary (or est. band), location, language requirement, key stack, must-have / nice-to-have.

Step 1B — find alt-source links

Skip this whole step if config/sources.json has "crossLink": false (the user wants only the links they gave — no alt-source lookup, no altLinks). Otherwise build an altLinks array of 1–4 entries on other platforms. Schema:

"altLinks": [
  {"platform": "Djinni",   "label": "Djinni",          "url": "https://djinni.co/jobs/..."},
  {"platform": "Company",  "label": "Company careers", "url": "https://example.com/careers/..."},
  {"platform": "LinkedIn", "label": "LinkedIn",        "url": "https://www.linkedin.com/jobs/view/..."}
]

Allowed platform values (drive colour-coding in the viewer): DOU · LinkedIn · Djinni · Company · Indeed · Other. Never duplicate the vacancy's own url. Never ship a 404 — verify with WebFetch before adding.

How to find them (in order):

  1. Djinni — WebSearch site:djinni.co "<company>" <role-keywords>. Match on company and ≥2 role-title tokens. Djinni's free-text search via plain fetch is unreliable; Google's site: search is the only thing that filters cleanly without Chrome.
  2. Company careers — WebSearch "<company>" careers <role>. Pick the result on the company's own domain; WebFetch to confirm the role is on the page (or accept a careers index if the specific posting URL isn't discoverable, and label it "Company careers").
  3. LinkedIn (optional) — WebSearch site:linkedin.com/jobs "<company>" <role>. Skip if nothing solid surfaces in 1–2 tries; never synthesise a URL from a job-id guess.

For an altLinks lookup loop across many vacancies, batch the WebSearch calls in parallel.

Step 2 — assess and score

Read profile/candidate.md first. Then for each vacancy, populate:

  • ss — five 0–100 sub-scores: tech, comp, level, company, appeal. See docs/rubric.md for what each axis means and how to anchor numbers.
  • score + tier — compute via:

    python3 scripts/score.py compute <tech> <comp> <level> <company> <appeal> [risk]
    # → score=N tier=X

    Don't eyeball. Don't hand-set tier. risk (optional, ±5) is for facts that don't fit an axis (visa friction, opaque domain). Clamped to ±5.

  • plus / minus — 2–4 bullets each, naming the real overlaps and gaps against the profile. Be honest about gaps (fine-tuning, 5+ yr bars, non-Python, recsys, etc.).
  • summary — one sentence: what they'd actually do day-to-day.
  • salary / salNum — as stated, or est. $X–Yk if not. salNum is the integer midpoint (for sorting).
  • loc — as stated.
  • voice — true only for voice/speech roles (these use the Voice CV).
  • cv — {variant, verdict, changes[], snippet}. variant: LLM (default) or VOICE. verdict: asis / light / real. changes: concrete CV edits. snippet: paste-ready text or null.
  • coverLetter — a tailored cover letter / application blurb, only if the user asks (or run /cover-letter later). The legacy field note is still read by the viewer.
  • status: default to_apply unless the user states otherwise (e.g. watch).
  • isNew: true.

Never infer a pipeline stage (hr_screen, rejected, etc.) from the posting — only from what the user tells you about their conversations.

Step 3 — preview

Compact preview, sorted by score desc:

NEW (N):
  - 86  S   appflame      AI Engineer                Remote
  -  ...
Skipped (already tracked):
  - <co> · <role>  →  existing id=<id>, status=<status>
Failed (manual follow-up):
  - <co> · <role>  →  <url>  (reason)

Ad-hoc runs: ask "Add these N roles?" Wait for approval. Bulk runs invoked from /refresh-listings operate with pre-approval — just append.

Step 4 — append, validate, build

data/vacancies.json is a v2 envelope ({schemaVersion: 2, vacancies: [...]}) — append new objects to its vacancies[] array (see scripts/board.py helpers). Keep the existing schema.

  • id — short kebab-case slug (appflame, or djinni-820717 for Djinni's numeric id). Don't duplicate an existing id; if the role already exists, update in place.
  • Position is cosmetic — the viewer re-sorts at render time. Append at the end for diff clarity.

Then always:

python3 scripts/validate.py        # exit 1 on schema issues
python3 scripts/build.py --real    # refresh the gitignored file:// snapshot (never the committed demo)

Skipping validate is a bug (schema drift creeps in silently). Skipping the build is a bug (your local file:// snapshot shows stale data; served api mode reloads on its own).

Edge cases

  • Djinni hides company names until shortlist — set co: "Hidden (Djinni)" and derive the role from the URL slug.
  • Aggregator URL — hand off to /refresh-listings instead. That skill has the per-source scraping playbooks (DOU "more" button, Djinni lazy-load, LinkedIn scroll container).
  • Salary "estimate" — only estimate if the JD doesn't state a band. Mark est. in the string.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.