06-reference

every ai powerpoint automation

2026-06-29·reference·source: Every·by Mike Taylor
ai-automationpresentation-toolingagent-skillsclaude-codeconsulting-workflows

Why this is in the vault

A practitioner-level account of why AI PowerPoint automation is harder than it looks, and how Every's consulting team built their own solution on top of Anthropic's official pptx skill.

The core argument

Every's consulting team (Mike Taylor, author) needed AI-assisted slide creation for roughly 2-3 decks per week. They started with Claude Code's PowerPoint generation but found it failed in two key ways: (1) it breaks down when working with existing company templates, because matching spatial layout, design aesthetics, and narrative structure simultaneously exceeds what a context-window-bound model can reliably do; and (2) .pptx XML is token-inefficient and was never designed for agents, so Claude can't reliably predict text wrapping or image placement.

The article frames the core failure mode as "80 percent right is often worse than zero" — a polished-looking deck with hidden errors is harder to catch than a manual one, because reviewers over-trust AI outputs. The team's senior applied AI engineer (Nityesh Agarwal) extended Anthropic's official pptx skill (59 files, 16 Python scripts, 4,000-word SKILL.md) and embedded it into their internal AI assistant "Claudie." The article teases (behind paywall) that a "blueprint-first" approach was the key breakthrough — generating a structural plan before touching the deck format. The author also flags that context rot above ~200K tokens degrades slide quality noticeably.

The conclusion is frank: most organizations shouldn't invest in PowerPoint automation yet — the engineering lift is outrageous relative to the current reliability ceiling.

Mapping against Ray Data Co

Strong reinforcement of two RDCO disciplines:

  1. Skill-file depth as a prerequisite for reliable automation — Anthropic's own pptx skill runs to 59 files and still fails at templates. This directly validates RDCO's "thin harness, fat skills" philosophy and the implementation-notes pattern. Any client engagement promising "AI will do your slides" should be scoped with the same rigor.

  2. 80%-right failure mode in consulting contexts — Taylor's observation that a near-miss deck is worse than a manual one maps directly to RDCO's verification-as-independent-worker pattern. For client-facing artifacts, the quality gate has to be near-zero-defect or the automation creates more review overhead than it saves.

Gap surfaced: RDCO doesn't have a documented approach to structured document generation (slides, reports) for client deliverables. As phData consulting work scales, this will become relevant — worth a spec if deck volume increases.

Related