job-search

Files (11)

• SKILL.md
• greenhouse-board.py · bin/
• hn-wih.py · bin/
• hn-wih.sh · bin/
• jd-pull.py · bin/
• linkedin-jobs.py · bin/
• pool-append.py · bin/
• remoteok.py · bin/
• remotive.py · bin/
• workingnomads.py · bin/
• test_parsers.py · tests/

SKILL.md (first lines)

# job-search

Scripted sourcing of remote job postings into a filtered, persistent pool.
This skill owns the mechanical parts: fetch, filter, render, append. Judgment
parts (fit assessment, tailoring, applying) are out of scope — hand a target
to the `job-application` skill once shortlisted.

## When to use

- User asks to find remote jobs in a market (e.g. Canada) from named boards or companies.
- A `_worklog` job-search program exists (or should be created) as the pool hub.

## Source ladder — cheapest first

1. **Public boards via HTTP** (no browser): `greenhouse-board.py` (with
   `--jd <id>` for full JD text), `hn-wih.py` (with `--max-age-days`),
   `remoteok.py`, `remotive.py`, `workingnomads.py`. Run these before any browser work. Himalayas.app has a paginated JSON API (`/jobs/api`, thin per page). Working Nomads yield skews agency-mediated (Lemon.io, Proxify) — flag those as staffing roles.
   **ATS board discovery**: probe `jobs.ashbyhq.com/<slug>` /
   `boards-api.greenhouse.io/v1/boards/<token>/jobs` / `api.lever.co/v0/postings/<slug>`
   across name variants — most companies live on one of the three (observed
   2026-09: groundnews, hopper, thumbtack, ascend, xero, kake, oscilar,
   clickup, gcai→gc-ai all on Ashby; discover slugs from careers-page embeds).
   **Board notes**: arbeitnow is EU-only; jobicy returns malformed responses;
   WeWorkRemotely hard-blocks HTTP (browser only); Job Bank Canada needs
   interactive search and skews below staff-level tech; Work at a Startup
   blocks plain HTTP.
2. **Direct posting URLs**: `jd-pull.py <url>` (shared with `job-application`).
   Exit 3 means JS-only shell — escalate that URL to the browser harness.
3. **LinkedIn via the browser harness** (background tabs only): run
   `linkedin-jobs.py` under `browser-use`. Two modes:
   `LJ_MODE=keyword` (keyword + location search) and `LJ_MODE=company`
   (a company slug's `/jobs/` page). LinkedIn's AI search may rewrite filter
   URLs and drop `f_WT=2` — capture the remote flag per posting from its
   location text, never from the URL. Per-card posting URLs are frequently
   absent from the results DOM in the AI-search variant (observed 2026-09:
   one `/jobs/view/` anchor page-wide); when `url` comes back empty, take
   company + title to the direct ATS ladder (greenhouse board tokens,
   ashby board JSON) instead of fighting the page.

## Invocation

```bash
# HTTP sources (stdlib only)
skills/job-search/bin/greenhouse-board.py instacart --country canada --remote-only
skills/job-search/bin/hn-wih.sh                     # latest Who-is-hiring month
skills/job-search/bin/remoteok.py --country canada

# LinkedIn (needs the browser-use daemon; BU_CDP_URL or local Chrome)
BU_CDP_URL=http://<host>:9223 browser-use < skills/job-search/bin/linkedin-jobs.py

# JD text for a direct URL (also: skills/job-application/bin/jd-pull.py)
skills/job-search/bin/jd-pull.py https://boards.greenhouse.io/example/jobs/123

# Persist rendered rows into the program's ## Pool section
skills/job-search/bin/pool-append.py --file people/oss/active/<program>.md \
  --source "Greenhouse instacart 2026-09-25" --rows-file rows.md
```

`linkedin-jobs.py` configuration (env): `LJ_MODE=keyword|company`,
`LJ_QUERY`, `LJ_LOCATION` (default Canada), `LJ_COMPANY` (slug),
`LJ_LIMIT` (default 8). It prints markdown rows; add `LJ_FORMAT=json` for rows
as JSON. It never activates a tab except for the one retry the harness skill
prescribes when a background scroll times out, and never types, clicks, or
applies.

## Discipline

- Read-only: fetch and extract only. No auto-apply, no form submission, no
  third-party posting. Login walls stop and ask the user.
- Posting facts rot: every persisted row carries its gather date; re-verify a
  link the day an application goes out (invariant echoed in the pool program).
- One tab per site/task; reuse a matching tab before opening a new one.

## Persistence

The pool lives in a `_worklog` program task (kind `program`, parent of any
per-target application tasks). `pool-append.py` inserts a dated `###` block
under `## Pool`; per-target tasks are created with
`"$WORKLOG_BIN/project.sh" add-child <program-slug> <child-slug> --kind=plan
--title=...` so the stub and the parent's `tasks:` entry land in one commit.
Commit subjects follow `_worklog/AGENTS.md`: `<slug>: <what changed>` with
`Worklog-Kind:` on creates.

From cheshirecode/dotfiles, Unlicense.

More from cheshirecode/dotfiles