AgriciDaniel/secretary ? reverse-engineered prompt

Build me an experimental AI secretary for Claude Code and Codex that helps with staff work, but stays strict about evidence and human approval.

I want it to take a task from a file, freeze the exact prompt and supporting notes before running, pull in only the relevant workspace evidence instead of everything, and send that to the chosen model. After the model responds, it should check the result locally against a strict schema, keep any contradictions or unknowns, and produce a clean decision ready summary with an evidence report. If the answer would lead to a real action like writing a file, sending something, spending money, or changing an account, it should stop and ask for explicit approval first.

Please include a simple command line flow for preflight, prepare, run, and result, plus install support for both Claude Code and Codex. Make it feel local first and careful, with profiles for different kinds of work like general, operations, communications, research, and chief of staff. If you need to look up current docs online, do that.