What problem does it solve? When an AI agent works with confidential data (PII, credentials, proprietary code), sending raw data to a remote model creates a leak risk. This Skill defines how to delegate sensitive work to a local subagent so the remote model only receives aggregated, sanitized results. ## Core Features & Use Cases - Delegation request template: A four-part structure for writing delegation prompts — concrete task, exact response format, field-by-field prohibition list, and explicit failure behavior. - Delegation boundary rules: Clear tables of what to delegate (aggregation, schema extraction, counting, redaction) and what not to delegate (long reasoning, multi-step chains, non-trivial code). - Local model selection criteria: Requires native tool-calling support, prefers the smallest model passing an acceptance test, and includes field observations on models like qwen2.5:7b and llama3.1:8b. - Use Case: You need counts of customers per plan from a CSV containing CPF numbers. The Skill instructs you to delegate the read-and-aggregate step to a local model with an explicit ban on returning names, CPFs, or emails, so only plano=<nome> qtd=<n> lines leave the machine. ## Quick Start Ask the agent to delegate reading a sensitive CSV to a local subagent, returning only per-plan counts with no personal identifiers in the response.
