Humanize Presentations — AI skill for Claude Code

humanize-presentations

**Remove AI tells from slide decks without destroying the ones that were deliberate.**

[](LICENSE) [](https://skills.sh/) [](#install)

A skill for AI coding agents that de-slops an existing `.pptx`. It handles what prose humanizers miss: deck-specific tells, the difference between slide text and speaker notes, non-English decks, and the file mechanics that make a repaired deck actually open.

---

Why this exists

Prose humanizers are excellent at prose. Point one at a slide deck and it does real damage, because its rhythm and voice passes tell you to add sentence fragments, first person, and lived detail. That advice is correct for a speaker's notes and wrong for a slide.

They also miss the tells that actually mark a deck as machine-made. A word list built for essays has nothing to say about topic-label titles, "Thank You / Questions?" slides, or a chart titled `Revenue by Quarter` instead of what the revenue did.

But the harder problem is the opposite one. **Most deck "tells" are also legitimate design choices.** A deck can follow a fixed section cycle because a rubric demanded it, and carry topic titles because the section names mirror an assignment. A checklist run without judgment flags those alongside the real tells, produces thirty findings, and the three that mattered get lost. The opposite failure is worse: a gate wide enough to excuse the rubric also excuses the template, and passes a deck with the section name in the eyebrow and the footer of every slide.

This skill exists to tell those two situations apart.

---

The core idea: two independent passes

Findings come in two classes, and conflating them is the failure mode this skill is built around.

| | Sentence-le