Code vs LLM judgement
The core rule: do not call AI where code suffices. Most of the pipeline is plain code at 0 tokens; the model is invoked only where judgement is needed, and cached by input hash.
What runs where
TaskRuns as
Parse pasted job, dedupecode · 0 tokens
Hard filter (location, visa, salary)code · 0 tokens
Legitimacy tier (green / yellow / red)code · 0 tokens
First-pass ATS check (fonts, links, formatting)code · 0 tokens
Overdue follow-up badges, analyticscode · 0 tokens
Recruiter rubric scoreAI · low
Tailoring, cover letter, pitchAI · on demand
Caching
Search is code, not AI. Code collects; AI evaluates. "Auto-search jobs by prompt" is a costly myth.