playbook-job-posting-language

Resolve each bare domain to the company's job-board profile, then query a jobs API with a date_posted window and description keyword filters applied server-side. The playbook returns one cleaned clause per row plus an evidence URL and posted date.

Most jobs APIs never return the full description, so scanning returned fields locally misses nearly everything (1 of 8 versus 5 of 8 measured). Pass keywords through the API's description filter, and check for silently stripped unknown keys in the request body.

Start with keywords and measure recall first: three API calls plus a 25-row sample. Use keywords only at 90% recall or above, add a gated AI classifier between 60-90%, and fix the keyword set below 60% because AI will not rescue bad keywords.

Detect postings over a 60-day window but suppress the line in copy when the posting is older than 30 days, using a formula rather than a model call. Bind copy only to the gated job_posting_line_safe field, never the raw ungated line.

A reasoning model at default effort burns the completion token budget on hidden reasoning, returning finish_reason length with empty content while abstaining rows still succeed. Set reasoning effort low with a 2500-token cap, or use a non-reasoning model like gpt-4o-mini inside Clay.

Do not use it to count hires or headcount growth over time, which is a different playbook, or to discover new companies hiring a role, which requires a market-wide search endpoint. It also cannot be a mandatory copy variable since roughly 60% of rows blank after the freshness gate.