Source pages were grouped into large text chunks and sent, through a batch interface, to the language model claude-opus-5-5 with instructions to extract every remote position it could identify. Remote meant remote, work-from-home, or hybrid with explicit remote allowance. Positions that were clearly on-site only were skipped, and pages without remote positions returned nothing.
Fields extracted
| Field | Description |
|---|---|
| Title | The position title as written |
| Company | The employer, if identifiable (missing for 128 deduplicated positions) |
| Role family | One of nine fixed families; see role families and seniority |
| Seniority | The level stated or clearly implied by the title, otherwise "unspecified" |
| Skills | Technical skills the posting asks for: languages, frameworks, databases, clouds, tools, methods and certifications, using common canonical names |
Scale
Extraction produced 4,512 raw positions. The run used about 3.67 million input tokens and 0.50 million output tokens. Output was constrained to a strict schema, and pages whose output overflowed were split and processed again rather than truncated.
A separate analysis step
After aggregation, the same model received the exact counts (not the raw pages) and produced the skill categories, the ten key insights, the ten ranked learning recommendations and the executive summary. Every claim in that analysis was required to come from the supplied numbers.