Self Presenting Decks

Self-Presenting Decks (orchestration)

Overview

Producing a deck that presents itself is a three-skill pipeline with a fixed order. This skill is the map — each stage's depth lives in its own skill.

deck (pptx) ──▶ narration (audio + autoplay pptx) ──▶ video (mp4)
   content            depends on final slides          depends on final slides + final audio

Order is the whole game: content → narration → video. Narration scripts describe what's on the slides ("look at the tree on the left"), so slides must be final first. The video embeds both the rendered slides and the audio, so it's always last. Editing upstream invalidates downstream — see the update matrix.

The stages

| Stage | Skill to use | Output | |---|---|---| | 0. Intake | — | ElevenLabs voice ID + ELEVENLABS_API_KEY (env var, never echoed); language/register choices (these live in narrating-pptx, not here) | | 1. Build/edit the deck | document-skills:pptx (pptxgenjs or template editing); presentation-generator for AI-image decks | deck.pptx (the clean copy — keep it) | | 2. Validate rendering | office-render — real PowerPoint → PDF → images | layout proof + the PDF the video will reuse — always from the CLEAN pptx (the narrated copy renders speaker icons onto slides) | | 3. Narrate | narrating-pptx — scripts → human approves scripts BEFORE TTS (it's real credit spend, often in a cloned voice) → TTS → embed → PowerPoint-authored autoplay → its own validation pass proves the narrated file clean | narration/scripts.json, narration/slideNN.mp3, deck-narrated.pptx | | 4. Video | deck-to-video — the stage-2 clean-deck PDF + the same narration/ mp3s → mp4 (progress bar + slide counter are defaults; flags turn them off) | deck.mp4 |

Keep narration/ and all artifacts as siblings next to the deck.

Three artifacts, three audiences

• Clean pptx — live presenting; humans talk over it.
• Narrated pptx — guided self-review at the viewer's own pace (they can linger, skip, re-listen per slide).
• mp4 — zero-friction async sharing; plays anywhere, but forces the narration's pace. Never ship only the mp4 to someone who needs to study the content.

Keep all three named as siblings (Deck.pptx, Deck_Narrated.pptx, Deck.mp4) so nobody presents the narrated copy live by accident.

The update matrix — what to rebuild when something changes

| What changed | Rebuild | |---|---| | Slide content/visuals | deck → re-render the whole PDF (PDFs aren't patchable per page) → re-check affected narration scripts (they reference what's visible) → regenerate only the changed clips → re-embed + autoplay → video | | Narration text only | the changed clips → re-embed + autoplay pptx → video | | Narration pace ("too slow") | ffmpeg -filter:a atempo=1.1 on the mp3s (pitch-preserving, no TTS cost) → re-embed → video | | Video overlay only (bar, counter) | video only — pptx artifacts untouched |

Regenerating only changed narration clips saves TTS cost — but always rebuild the narrated pptx and video from the full final audio set.

Hard-won rules that hold across the whole pipeline

• Never hand-write PowerPoint `` autoplay XML — corrupts the file. PowerPoint authors it via the narrating-pptx script. A video-only request is the escape hatch when PowerPoint isn't available.
• All slide rasterization comes from real PowerPoint (office-render), never LibreOffice — fonts and layout re-flow otherwise, and the video won't match the deck.
• One narration source of truth: the pptx and the video must embed the same mp3s, same pace — otherwise reviewers hear two different presentations.
• Ear-testing is human work: autoplay start and audio levels can't be verified headlessly — say so explicitly when handing off.

Choosing the entry point

• "Add narration/voiceover to my deck" → stages 2–3 (+ offer 4; the mp3s make the video nearly free).
• "Make a video of my deck" with no narration yet → the full chain; the narration is where the work is.
• "Make a video" and narration already exists → stage 4 only, reusing narration/.
• Deck doesn't exist yet → stage 1 first; don't write narration for unbuilt slides.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: shaharsha
• Source: shaharsha/claude-skills
• License: MIT

Install and usage instructions live in the source repository linked above.