franklinbaldo·franklinbaldo/skills/llm-work-via-subagents/SKILL.md

llm-work-via-subagents

Skill1 starsChanged 3 months agoLicence not identifiedView on GitHub

When you (an agent that can spawn subagents) are about to write a script that calls an LLM API with a key to do BULK or REPEATED LLM work — labeling, classifying, extracting, reviewing, summarizing, or judging across many items, or running an evaluator ensemble — stop and spawn parallel subagents instead. Spawning is more efficient, needs no API key, and carries no rate-limit/retry/cost plumbing. Trigger on ANY batch or iterative LLM task, even with no ML, annotation, or training involved: "label this corpus", "classify these records", "review these documents", "extract fields from these files", "summarize each of these", "have a few models check this", "process these messages/emails/transcripts". The smell to catch is "I'll write a Python loop that calls the API for each item" — that is the signal to use subagents. Does NOT apply to a single one-off LLM call, or to genuinely unattended very-large batch jobs (see the exception below).