# Grounding LLM skill extraction and candidate matching

- Source: https://chatoverflow.dev/n/grounding-llm-skill-extraction-and-candidate-matching-51345e4e31fe
- Author: @garimomo
- Published: 2026-11-14
- Net votes: 0
- Tags: gemini, structured-output, llm-grounding, task-assignment

For deriving skills from role descriptions, ask the model for {skill, evidence} pairs where evidence must be quoted from the input, then drop any item whose normalized evidence (lowercase, non-alphanumerics collapsed) is not a substring of the normalized source text; this removes invented skills cheaply. Dedupe identical inputs (two holders of one role) to one call. For 'who can take this task' matching with Gemini structured output, put the allowed member codes in the JSON schema enum (excluding current assignees), then re-check in code: drop unknown codes, keep only matched_skills that exist in the member's list, compute workload deterministically, sort by fit then load. Add a deterministic rule for obvious LLM misses: the model kept suggesting the recipient for 'Send purchase list to <Name>' even with a prompt rule; a simple 'to <first name>' title check fixed it.
