The body is lean — an ASCII pipeline, compact tables, concrete code blocks, and a distilled Gotchas list — and assumes Claude's competence without explaining concepts Claude already knows, fitting the 'every token earns its place' anchor.

Provides fully executable, copy-paste-ready guidance: ffmpeg commands, an edge-tts Python snippet, soffice/pdftoppm conversion, and a Dependencies install table, rather than the pseudocode or vague direction of lower anchors.

A four-stage pipeline is laid out with an ASCII diagram and numbered stages, reinforced by an explicit 'All segments MUST share: 44100Hz, stereo, AAC' verification checklist and a Gotchas section covering error recovery for the batch concat operation.

The body is an overview that points one level deep to real, clearly signaled references — slide-generation.md, tts-engines.md, pptx-conversion.md — and batch scripts (all verified to exist), matching the well-signaled one-level-deep anchor.

Lists multiple concrete actions — 'Create academic presentation slide decks and optionally demo videos from research papers' and 'Covers slide generation, narration scripts, TTS audio, and video assembly' — matching the top anchor rather than the single-action level 2.

Explicitly answers both what (decks/videos from papers; generation, narration, TTS, assembly) and when via a clear 'Use when the user asks...' clause, so it is not capped at 2 for a missing trigger clause.

Includes a dense set of phrases users naturally say — 'make slides', 'create a deck', 'make a presentation', 'demo video', 'paper slides', 'conference talk slides' — giving good coverage beyond the partial level 2.

Targets a clear niche (research papers → academic presentations) with distinctive triggers, making overlap with other skills unlikely; it is more specific than the level 2 'somewhat specific' anchor.