Present Agent Skill

Present — Narrated Interactive Presentations

Generate a self-contained HTML presentation with dual article/slides mode, ElevenLabs narration, optional GPT Image 2 illustrations, and scroll-reveal animations.

What This Skill Produces

A single index.html file (plus audio and optional image assets) that can be:

• Opened locally in a browser
• Deployed to Vercel, Netlify, or any static host
• Shared as a folder

The output has two modes the viewer can toggle between:

• Article mode — long-form scrollable report with Tufte-inspired typography
• Slides mode — navigable presentation with keyboard/click navigation and narrated audio playback

Quick Start

/present "AI adoption research for Arseny" --slides 12 --voice daniel --images risograph

Or with a file:

/present path/to/research.md --detail detailed --voice alice

Parameters

| Parameter | Values | Default | Description |

|-----------|--------|---------|-------------|

| --slides | 5-20 | 12 | Number of slides |

| --detail | executive, standard, detailed | standard | Content depth |

| --voice | ElevenLabs voice name | daniel | Narrator voice |

| --images | style name or none | none | Image generation style |

| --image-prompt | custom string | auto | Override image prompt prefix |

| --output | path | ./presentation/ | Output directory |

| --deploy | vercel project or none | none | Auto-deploy target |

| --title | string | auto | Presentation title |

| --no-audio | flag | false | Skip audio generation |

Detail Levels

• executive (5-7 slides): Key findings only. One stat slide, one recommendation slide, sources. Best for busy stakeholders who need the bottom line.
• standard (10-14 slides): Full narrative arc. Problem, evidence, analysis, recommendations, sources. The default for most presentations.
• detailed (15-20 slides): Deep dive. Includes methodology, multiple evidence sections, case studies, detailed recommendations with implementation steps.

Voice Options

Uses ElevenLabs API. The key must be available in ~/claude-skills/elevenlabs-tts/.env as ELEVENLABS_API_KEY.

Recommended voices for presentations:

• daniel — Steady Broadcaster, British, formal (default)
• alice — Clear Educator, British, professional
• matilda — Knowledgeable, American, upbeat
• brian — Deep Resonant, American, comforting
• george — Warm Storyteller, British, mature

Image Styles

When --images is set, the skill generates illustrations for key slides using GPT Image 2 (~/.claude/skills/gpt-image-2/scripts/gpt_image_2.py). Available styles:

• risograph — Gerd Arntz isotype style, muted colors, sand texture
• editorial — Magazine photography style, dramatic lighting
• blueprint — Technical drawing aesthetic, white on blue
• ink — Black ink illustration, hand-drawn feel
• constellation — Data visualization aesthetic, dots and lines
• Custom: pass --image-prompt "your style description" to override

Images are generated in --draft mode first (~$0.006/image). The skill decides which slides benefit from illustration (typically 3-5 out of 12).

Workflow

Step 1: Content Analysis

Read the input content (a topic description, a markdown file, vault notes, meeting transcript, or research). Identify:

• The core argument or narrative
• Key data points and statistics
• Natural section breaks
• Quotable findings with sources

Step 2: Slide Planning

Based on --detail and --slides, create a slide plan. Each slide needs:

Slide N: [Type] — [Title]
Content: [what appears on screen]
Narration: [what the voice says — always more than what's on screen]
Read time: [seconds for an average reader to absorb the visual content]
Image: [yes/no, with prompt if yes]

Slide types: title, summary, stat, evidence, comparison, quote, framework, recommendation, case-study, sources

The narration script should be conversational and add context beyond what's displayed. It should NOT just read the slide text aloud — it should explain, connect, and elaborate. Target 15-30 seconds of narration per slide.

Step 3: Generate Audio

For each slide, generate narration using ElevenLabs:

python3 ~/.claude/skills/elevenlabs-tts/scripts/elevenlabs_tts.py \
  --voice <voice_name> \
  --text "<narration>" \
  --output <output_dir>/audio/slide-<N>.mp3

Or use the direct API via the script at scripts/generate_audio.py in this skill.

Also generate a transition sound (Rhodes chord) for slide-to-slide transitions.

After generation, get durations with ffprobe to calculate slide timing.

Step 4: Generate Images (if enabled)

For slides that benefit from illustration, generate images using GPT Image 2:

python3 ~/.claude/skills/gpt-image-2/scripts/gpt_image_2.py --draft --size 1536x1024 \
  "<style prefix> <slide-specific prompt>" \
  <output_dir>/images/<name>.png

Typically generate 3-5 images for a 12-slide deck. Choose slides where a visual metaphor strengthens the point — stat slides, concept slides, and the title slide are good candidates. Don't illustrate every slide.

Step 5: Build HTML

Use the template at assets/template.html as the base. The template includes:

• Typography: EB Garamond (body) + DM Sans (labels/numbers)
• Color palette: Configurable via CSS variables in :root
• Article mode: Tufte-inspired layout with executive summary box, stat cards, two-column sections, data tables
• Slides mode: Full-viewport slides with fade transitions, keyboard navigation (arrows, space), dot indicators
• Audio engine: Single reusable <audio> element, slide-synced playback with progress bar, transition sounds between slides
• Auto-hide controls: Top bar (mode switcher + audio) appears when cursor enters top 20% of viewport. Bottom nav appears in bottom 20%. Shift+. toggles always-show/always-hide/zone mode.
• Scroll-reveal animations: Intersection Observer-based fade-up for sections, staggered stat cards, animated counters, h2 rule-draw effect
• prefers-reduced-motion: All animations disabled when user prefers reduced motion

Populate the template by replacing placeholder sections with the actual slide and article content.

Step 6: Test

Open in browser using /real-browser or open <path>. Verify:

• [ ] Article mode renders correctly, images load
• [ ] Slides mode: all slides navigable, text fits within viewport
• [ ] Audio plays when play button is clicked
• [ ] Audio syncs to slide advancement (each slide waits for narration + read time)
• [ ] Transition sounds play between slides
• [ ] Auto-hide works for top and bottom bars
• [ ] Keyboard navigation (arrows, space) works in slide mode

Step 7: Deploy (if requested)

If --deploy is set, copy output to the target project's public/ folder and deploy:

cp -r <output_dir>/* <project_path>/public/<slug>/
cd <project_path> && vercel deploy --prod --yes

HTML Architecture

Audio Sync Model

Each slide has three timing properties:

• data-audio="slide-name" — maps to audio file
• data-read-time="N" — seconds for reading the visual content

The audio engine calculates: slide_duration = max(audio_duration, read_time) + 2s. After narration ends, it waits for any remaining read time plus a 2-second buffer, plays a transition sound (1.8s), then advances to the next slide.

Avoiding AI-Looking Formatting

The following patterns read as AI-generated and should be avoided:

• Colored left-bar + bold heading + description blocks (finding cards)
• Large italic pull quotes with colored left border
• Uniform card grids with icon + heading + description
• Gradient text on metrics

Instead use:

• Natural prose paragraphs with inline emphasis
• Definition lists (<dl>) for structured points
• Tables for comparisons
• Direct statements woven into flowing text

Image Paths

Use absolute paths from the deployment root: /slug/images/name.png, not relative paths. Relative paths break when URLs load without trailing slashes.

Files

• SKILL.md — This file
• scripts/generate_audio.py — ElevenLabs TTS batch generator
• assets/template.html — Base HTML template with all CSS/JS
• references/slide-types.md — Detailed slide type specifications and examples