Every listing gets an AI-Agency Score (0–100) and two employer-openness flags. This is the rubric, transparently.
Claude reads every job description, scores two axes (Build + Use), computes a blended AI-Agency Score, and flags whether the employer affirmatively welcomes AI-assisted workflows and AI tools in interviews. The split mirrors the Lightcast Open Skills taxonomy dichotomy between "AI engineering" (builds AI) and "AI literacy" (uses AI).
Does this role BUILD AI systems?
High scores for ML research, LLM infra, training/inference engineering, foundation model work, agentic architecture. Low for sales / ops roles that barely touch AI.
Does this role USE AI tools day-to-day?
High scores for roles explicitly centered on Claude / Cursor / Copilot / agentic workflows. Low for legacy / process-heavy roles with minimal AI exposure.
Build is weighted heavier because in 2026 "using AI tools" is becoming universal — it's a weaker differentiator. "Building AI" is still the rarer, more distinctive skill. That said, we think both matter, so we don't zero out Use. The 60/40 split lets a high-Use / low-Build candidate still land in AI-fluent (e.g. a growth marketer who lives in Claude + n8n scoring ~65).
Default public view shows only AI-fluent (50+) and AI-core (75+). Jobs scoring 20–49 ("AI-touching") still exist at /ai-touching or via ?min_agency=20, but aren't promoted — they're tech jobs adjacent to AI, not AI jobs, and the site's promise is the latter.
Research surfaced a critical nuance: on-the-job AI policy and interview AI policy diverge in practice. Anthropic welcomes AI on the job but says "applicants should not use AI assistants". Amazon forbids AI in interviews but tolerates on the job. A single "AI-friendly" boolean would mislead users. So we extract two flags:
round(0.6 × build + 0.4 × use), we override with the computed value. The LLM does subjective axis scoring; we do the arithmetic.The Build-vs-Use split is not original to us — it mirrors the dichotomy in the Lightcast Open Skills taxonomy (the de facto standard used by LinkedIn, Indeed, and most ATS vendors) which separates AI engineering (builds AI) from AI literacy (uses AI). Our contribution is applying it per-listing, with a numeric score, which is novel at the consumer-facing level.
Academic frameworks that informed this design:
No existing job board produces a per-listing 0–100 AI score. AIOE/SML/OECD all score at the occupation or sector level. Consumer-facing boards (aijobs.net, aijobs.ai, Wellfound) operate as binary "AI bucket" filters without granularity. Our contribution: applying the industry-standard Build/Use dichotomy to individual listings, transparently, using an LLM with a published rubric.
Scoring is done by Claude through a structured extraction pipeline. One merged prompt extracts all structured fields, categorization, both scores, and both employer-openness flags in a single call. Content-hash deduplication ensures unchanged jobs skip the LLM entirely on re-runs.
Email hello@ai-jobs.careers with the job URL and your reasoning. We re-score on a rolling basis and refine the rubric as the market evolves.