job-search-agent
Installation
SKILL.md
Job Search Agent
You are working on the agent itself — configuring, extending, or troubleshooting it — not running a search. Daily use lives elsewhere: onboarding and the home view in job-search, the search pass in job-search-run. This manual holds the playbooks and guardrails for changing the system safely.
The system in one paragraph: a private, local-first job-search agent. It searches LinkedIn, Ashby, Greenhouse, and Lever company-board postings via agent-data, judges each one against the user's prose preferences brief, and writes digests to a workspace that never touches source control (~/.job-search/ by default).
Hold these stances in every change you make — each exists for a reason, and several are CI-enforced:
- Qualitative-by-default. Relevance is expressed as relevant or not, and if relevant as weak / moderate / strong with plain-language reasoning. No score is computed or stored unless you explicitly ask — and even then a numeric score is never written into your saved digests,
jobs.jsonl, or preferences brief. - Frequency, in human terms. You tune the system by how often to pull: hourly, daily, weekly.
- Private-local. All data lives under
~/.job-search/(hidden, deny-all.gitignore). Nothing is ever committed to a public repo by the agent. - Every blocked path is a named error. If anything can't proceed, the agent names the exact
E-*with its cause and fix, then stops. No silent failures. - Conversational-first config. You change anything — a query, the frequency, your preferences — by chatting. The agent edits
config.yamlminimally and writes it back. Hand-editing files is an escape hatch you can always use, not a required step.