Your Board Deck Has a Wrong Formula. Excel Won't Flag It.

TL;DR

• Pretty files can be structurally wrong — Nate opens with a financial model that looked validated but had an incorrect revenue growth formula copied across future years, proving that clean formatting and plausible outputs are not the same as truth.

• The real shift is from prompts to workflows — His core claim is that a prompt asks for an output, while a workflow defines the stages an Excel file or PowerPoint deck must survive before it can be trusted.

• Serious AI document work has four stages — He lays out a repeatable system: prepare sources, define structure, build the artifact against that spec, and then verify it aggressively before sharing.

• He uses model-vs-model review loops in production — Nate says he often has Codex build Excel models or deck logic, then uses Claude Opus 4.7 as a hostile reviewer that generates edit lists he feeds back into Codex until the work reaches "A-level."

• Not all document tasks carry the same risk — He rates formatting, chart drafts, and wording as low risk, source attribution and extraction as medium risk, and numerical synthesis, financial calculations, regulatory language, and exec-facing claims as high risk.

• Why there’s no push-button solution — His answer is that knowledge work is too contingent on domain-specific context to be fully generalized, comparing the need to build your own workflow harness to Luke Skywalker making his own lightsaber.

The Breakdown