refresh-listings
Skill1 starsChanged 5 months agoMITView on GitHub
Use whenever the user wants fresh AI/ML roles pulled into Job Search HQ. Triggers: /refresh-listings, "refresh listings", "refill", "what's new", "scan for new AI jobs". Walks three sources in Claude in Chrome — DOU AI/ML category pages (no auth), the user's Djinni dashboard (auth required), and the user's LinkedIn recommended feed (auth required) — dedupes against `data/vacancies.json`, and hands each survivor to the /add-vacancy flow.
In this repository9 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- add-sourceSkill
- add-vacancySkill
- build-profileSkill
- cover-letterSkill
- ingest-cvSkill
- interview-prepSkill
- recheck-statusesSkill
- update-statusSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 3 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 144
Read profile/candidate.md and score via scripts/score.py compute.
- Distinctive: line 175
Stop clicking "more" / scrolling: when the count stops growing, not after a fixed number of iterations.
- Distinctive: line 177
Use path-only URLs: when extracting from the DOM — full URLs with long query strings can trigger a safety filter when fed back to other …
- Distinctive: line 179
Don't re-score existing entries: If a row already exists, leave it alone — even if its title in the listing has been edited slightly since …
≈ “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.
Files
refresh-listings
Pull fresh AI/ML roles from three sources and ingest the new ones.
Sources
Sources live in config/sources.json — walk every
entry with "enabled": true, in order, using its urls[]. Edit that file to add,
remove, disable, or repoint a source; no skill change needed. Selectors stay
here (sites rotate class names); config holds only the stable bits. The table
below mirrors the shipped defaults.
| Source | URL | Auth? |
| DOU | https://jobs.dou.ua/vacancies/?category=AI%2FML&exp=1-3 | no |
| DOU | https://jobs.dou.ua/vacancies/?category=AI%2FML&exp=3-5 | no |
| Djinni | https://djinni.co/my/dashboard/ (personalised recommendations) | yes |
https://www.linkedin.com/jobs/collections/recommended (the user's recommended feed) | yes |
Djinni dashboard and LinkedIn require the user's logged-in session — open them in Claude in Chrome. If the user isn't logged in (the landing page is the public sign-in screen), stop and ask them to log in before retrying. Don't try to scrape an anonymous version — the data is different and the dedup keys won't line up.
Flow (per source)
- Open the URL in a Claude-in-Chrome tab. Create a fresh tab group if needed.
- Expand the list:
- DOU: click
.more-btnuntil the listing count stops growing (typically 4–6 clicks). - Djinni / LinkedIn: scroll the inner list container until the count plateaus.
- DOU: click
- Extract
{title, company, url, salary, loc}for every visible row. Use the per-source selectors below. - Drop noise via the title-filter regex.
- Dedupe against
data/vacancies.json(path-match + company/title match). - Hand survivors to
/add-vacancyper role.
Cross-source dedup: a role on Djinni and the same role on DOU should produce one card, not two.
Match on (normalised company, normalised title).
Per-source playbooks
The selectors below are what work today. If a query returns zero rows, fall back to the generic anchor patterns suggested in the inline comments — sites rotate class names.
DOU
// Listing selector: li.l-vacancy
// "Load more" button: .more-btn a (falls back to .more-btn itself)
// Click pattern: click → wait 3s → re-count → repeat until stable.
JSON.stringify([...document.querySelectorAll('li.l-vacancy')].map(li => {
const a = li.querySelector('a.vt');
return {
title: a?.textContent.trim() ?? '',
path: a ? new URL(a.href).pathname : '', // path-only — long query strings can trip safety filters
company: li.querySelector('a.company')?.textContent.trim() ?? '',
salary: li.querySelector('.salary')?.textContent.trim() ?? '',
city: li.querySelector('.cities')?.textContent.trim() ?? ''
};
}));To reconstruct a full URL later: 'https://jobs.dou.ua' + row.path.
Djinni dashboard (/my/dashboard/)
// Cards under the dashboard render as job-list-item rows.
// Lazy-load: scroll the document until the count stops growing.
JSON.stringify([...document.querySelectorAll('li.list-jobs__item, .job-list-item')].map(li => {
const a = li.querySelector('a.job-list-item__link, a.h3');
return {
title: a?.textContent.trim() ?? '',
url: a ? new URL(a.href, 'https://djinni.co').toString() : '',
company: li.querySelector('.job-list-item__company-name, .text-body')?.textContent.trim() ?? '',
salary: li.querySelector('.public-salary-item, .job-list-item__salary')?.textContent.trim() ?? '',
loc: li.querySelector('.location-text, .nobr')?.textContent.trim() ?? ''
};
}));Fallback if empty: document.querySelectorAll('a[href^="/jobs/"]') and walk up to the
containing card.
LinkedIn recommended (/jobs/collections/recommended)
// LinkedIn renders the feed as cards inside a scrollable inner container.
// Scroll the *inner* container (not window) until the count stops growing.
JSON.stringify([...document.querySelectorAll('li.jobs-search-results__list-item, li.scaffold-layout__list-item')].map(li => {
const a = li.querySelector('a.job-card-list__title, a.job-card-container__link');
return {
title: a?.textContent.trim() ?? '',
url: a ? new URL(a.href, 'https://www.linkedin.com').toString().split('?')[0] : '',
company: li.querySelector('.job-card-container__primary-description, .artdeco-entity-lockup__subtitle')?.textContent.trim() ?? '',
loc: li.querySelector('.job-card-container__metadata-item')?.textContent.trim() ?? ''
};
}));LinkedIn rotates class names regularly. Fallback: document.querySelectorAll('a[href*="/jobs/view/"]')
and walk up to the parent <li>.
Title-noise filter
The canonical list is the titleNoise regex in config/sources.json
(case-insensitive) — edit it there to tune filtering. The shipped default:
/(Creator|Animator|Motion|Designer|SMM|Marketing|Teacher|Labeling|Product Manager|
Project Manager|Business Analyst|\bAnalyst\b|Recruiter|Drone|Embedded|Website|
Copywriter|Content|Computer Vision|\bCV Engineer\b|\bVision\b|Operator|artist|
Art Director|Robotics|UAV|SEO|Sourcer|Lead Generation|Vibe|No-code|Sales|QA|
Manager|Director|Annotator|Data Entry|Solutions Architect|Trainer|Product Owner|
Lawyer|Consultant|Sales Engineer|Solutions Engineer)/iDedupe rules
A new row is new only if all of:
- Its URL path isn't already a primary
urlindata/vacancies.json. (company, normalised role title)doesn't match an existing entry with a different primary URL — covers the case where the same role is tracked via the company's own site, or via a different aggregator.- Borderline cases (same company, near-duplicate title) → flag for the user, don't auto-add.
Hand-off to /add-vacancy
For each survivor, run /add-vacancy Step 1A onward:
- Fetch the JD.
- Find
altLinks. - Read
profile/candidate.mdand score viascripts/score.py compute. - Build the full entry.
status: "to_apply",isNew: true.
Sort the preview by score desc before confirming with the user. Bulk runs (the user invoked
/refresh-listings) operate with pre-approval — go ahead and append.
Finish
python3 scripts/validate.py
python3 scripts/build.py --real # refresh the gitignored file:// snapshot (never the committed demo)Output
Source: DOU 1-3 → kept N, deduped M, added K
Source: DOU 3-5 → kept N, deduped M, added K
Source: Djinni → kept N, deduped M, added K
Source: LinkedIn → kept N, deduped M, added K
Added (sorted by score):
- 86 S appflame AI Engineer Remote
- ...
Borderline (manual check):
- <co> · <similar-title> vs existing <id>Edge cases
- Stop clicking "more" / scrolling when the count stops growing, not after a fixed number of iterations. Some categories have 30 entries; others have 200+.
- Use path-only URLs when extracting from the DOM — full URLs with long query strings can trigger a safety filter when fed back to other tools.
- Don't re-score existing entries. If a row already exists, leave it alone — even if its title in the listing has been edited slightly since the user added it.
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
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.

