Descriptions:
Nate B Jones outlines a four-stage agentic workflow for producing reliable AI-generated Office documents — PowerPoint decks and Excel workbooks — using Codex for artifact creation and Claude Opus 4.7 as an adversarial reviewer. The talk is grounded in a real failure he encountered: a financial model that looked correct, passed a written validation check, but had a broken revenue formula silently copied across every projected year. That experience motivates his central argument: the gap between a prompt that produces a plausible-looking file and a workflow that produces a trustworthy one is what most teams are not yet closing.
The four stages he defines are source preparation (creating an indexed, annotated work packet before any file is built), structure (producing a file specification — narrative spine, slide list with claim headlines, or tab architecture and calculation flow — before any content is generated), artifact creation (building the file against the spec and source packet rather than from the model’s internal assumptions), and verification (using a second model, in this case Claude Opus 4.7, to aggressively review the finished file for errors, unsupported claims, and formula logic).
Jones also shares that he drafted eight simultaneous documents in a single week using this approach, and argues that treating agents as the center of the workflow rather than a bolt-on tool is what unlocks order-of-magnitude productivity gains for serious knowledge work. A source packet template including an ID schema and conflict log format is available on his Substack.
📺 Source: AI News & Strategy Daily | Nate B Jones · Published May 27, 2026
🏷️ Format: Workflow Case Study
