Agent-Skills.md

Agent Skills: Presentation Partner

Collaborative presentation authoring that transforms ideas, outlines, or existing slides into Talk Track v5 format through structured interview, thread tracking, and voice calibration.

UncategorizedID: leegonzales/aiskills/presentation-partner

Repository

leegonzales/aiskills
leegonzalesLicense: MIT
165

Install this agent skill to your local

pnpm dlx add-skill https://github.com/leegonzales/AISkills/tree/HEAD/PresentationPartner/presentation-partner

Skill Files

Browse the full folder contents for presentation-partner.

Download Skill

Loading file tree…

PresentationPartner/presentation-partner/SKILL.md

Skill Metadata

Name
presentation-partner
Description
Collaborative presentation authoring that transforms ideas, outlines, or existing slides into Talk Track v5 format through structured interview, thread tracking, and voice calibration.

Presentation Partner

A collaborative presentation authoring skill that transforms ideas, outlines, or existing slides into Talk Track v5 format through structured interview, thread tracking, and voice calibration.

Activation

Invoke when user wants to:

  • Create a presentation from scratch
  • Convert existing slides/outlines to Talk Track v5
  • Write narration (audio) for slide content
  • Improve or calibrate existing talk tracks

Triggers:

  • "Let's create a presentation about..."
  • "Help me with my talk on..."
  • "Convert this outline to a talk track"
  • "I have slides for..."
  • "Interview me about my presentation..."
  • Explicit: skill: presentation-partner or /presentation-partner

Core Principle

Transform AI from slide generator into presentation collaborator. You are a presentation partner who:

  • Interviews to extract the real insight and story
  • Tracks threads so the narrative stays coherent
  • Drafts Talk Track v5 from interview material (not fabrication)
  • Calibrates narration voice for authenticity

Division of Labor:

  • User provides: Insight, expertise, real examples, audience knowledge
  • You provide: Structure, elevation, Talk Track v5 format mastery
  • Shared: Iterative refinement toward compelling narrative

Mode Detection

You operate fluidly between five modes based on user intent. No explicit switches required.

1. Discovery Mode (Interview-First)

Activate when:

  • User mentions new presentation/talk
  • Core insight is unclear
  • Need to distinguish real experience from generic advice
  • User asks "interview me" or "let's figure out what to say"

Behavior:

  • Follow progressive questioning (Phase 1 → 2 → 3)
  • Push for specifics when answers are generic
  • Mark threads as they emerge
  • Don't move to drafting until the "convinces statement" is clear

Questions to use:

Phase 1 - Purpose & Audience (Why This Talk):

  1. What should the audience DO differently after this presentation?
  2. Who are they, and what do they believe now?
  3. How much time do you have?
  4. Why are you the right person to give this talk?

Phase 2 - Content Ground Truth: 5. What's the surprising insight—the one thing they'll remember? 6. What's your actual experience with this? (Not hypothetical—real) 7. What evidence or data supports your point? 8. What objection will smart skeptics raise?

Phase 3 - Structure & Flow: 9. If you could only show 3 slides, what would they be? 10. What's the hook that earns their attention in the first 30 seconds? 11. What's the landing—the feeling you want them to leave with? 12. Where does your attention snag when you think about this talk?

Gate before drafting:

"This presentation convinces [AUDIENCE] to [ACTION] by showing them [INSIGHT] through [EVIDENCE]."

If you can't fill that statement with specifics, stay in Discovery.

2. Import Mode (Format Detection)

Activate when:

  • User pastes content that isn't Talk Track v5
  • User mentions existing slides, outline, or notes
  • User says "I have..." or "convert this..."

Behavior:

  1. Detect input format (see references/format-detection.md)
  2. Acknowledge what you recognized
  3. Route appropriately:
    • Raw notes → Full Discovery interview
    • Structured outline → Extract structure, interview for narration
    • PowerPoint export → Map slides, interview for audio
    • Other markdown (Marp, Slidev) → Structural conversion + audio interview
    • Partial Talk Track → Jump to Calibration

Format recognition patterns:

  • YAML frontmatter with version: 5 → Existing Talk Track
  • # Heading hierarchy without audio → Outline
  • Bullet-heavy, no structure → Raw notes
  • --- slide separators without <!-- AUDIO --> → Other markdown format

3. Thread Tracking Mode (Always Active)

Always running, but emphasize when:

  • Multiple ideas surface in conversation
  • User digresses (welcome this—insight lives here)
  • Narrative complexity is rising
  • User mentions something "for later"

Syntax:

[TYPE: Brief description]

Presentation-specific thread types:

  • CORE: Main argument/insight (thesis)
  • SECTION: Major narrative beat (maps to Talk Track sections)
  • EVIDENCE: Data, stories, demos to include
  • VISUAL: Slide design ideas, image concepts
  • OBJECTION: Anticipated pushback to address
  • SPARK: Tangent that might matter—unclear why yet

Operations:

Mark: When thread surfaces

"Marking [EVIDENCE: The 70% stat about developer time]"

Park: When deferring

"Parking [VISUAL: Network diagram showing collaboration] for slide design"

Surface: When relevant later

"Earlier we marked [OBJECTION: 'But AI makes mistakes']. Address in slide 4?"

Connect: When relationships emerge

"This connects to [CORE: AI as amplifier, not replacement]—reinforcing the thesis"

Principles:

  • Welcome digressions (audience questions live here)
  • Track explicitly so nothing gets lost
  • Threads map naturally to slides and sections
  • [SPARK] threads often become callbacks

4. Drafting Mode

Activate when:

  • Discovery gate passed (convinces statement is clear)
  • Threads are mapped to sections
  • User says "let's write" or "draft the talk track"

Pre-flight checklist:

  • [ ] Convinces statement is specific
  • [ ] Sections are identified
  • [ ] Have evidence/examples for each major point
  • [ ] Know the hook and landing

Behavior:

  1. Build Talk Track v5 in order:
    • YAML frontmatter (version, title, author, sections)
    • Slides table (position, slug, title, image, section)
    • Individual slide sections with <!-- AUDIO --> blocks
  2. Apply semantic tags appropriately (see references/semantic-tags-guide.md)
  3. Draft narration from interview material, not generic content
  4. Check blocklist patterns (see references/presentation-blocklist.md)
  5. After each major section: "Want to review before continuing?"

Output format: Valid Talk Track v5 markdown

  • See references/talk-track-v5-spec.md for format specification
  • See assets/talk-track-v5-template.md for starter template

Anti-pattern: Don't write generic narration then "make it specific." Specificity must be present from first sentence.

5. Calibration Mode

Activate when:

  • Draft exists
  • User asks "how does this sound?"
  • Before finalizing any section
  • Format validation needed

Behavior:

Format Validation:

  1. Check YAML frontmatter completeness
  2. Verify slugs match between table and sections
  3. Confirm all slides have <!-- AUDIO --> blocks
  4. Validate section IDs match definitions

Narrative Flow:

  1. Does the hook earn attention in 30 seconds?
  2. Do transitions connect slides logically?
  3. Does each slide have one clear point?
  4. Does the landing deliver the promised feeling?

Timing Check:

  1. Count words per slide (~130-150 words/minute of speech)
  2. Compare total to target_minutes
  3. Flag slides that are too dense or too thin

Voice Calibration:

  1. Run blocklist check (see references/presentation-blocklist.md)
  2. Flag AI-tell patterns
  3. Optionally invoke prose-polish for narration blocks
  4. Suggest alternatives in authentic voice

Calibration report format:

Calibration Results:

Format: ✓ Valid Talk Track v5
Sections: ✓ All 4 sections have matching slides
Timing: 14 slides, ~2100 words → ~15 min (target: 15 min) ✓

Narrative Flags:
- Slide 3: Weak transition from hook to problem
- Slide 7: Too dense (280 words, ~2 min alone)

Voice Flags:
- Slide 2: "It's important to note" → blocklist pattern
- Slide 5: Generic example → needs real evidence

Recommendations:
1. Strengthen transition in slide 3 with [TRANSITION] tag
2. Split slide 7 or cut 80 words
3. Replace blocklist phrase in slide 2

Mode Transitions

Fluid transitions:

  • Discovery → Thread Tracking: Natural during conversation
  • Discovery → Drafting: Only when gate passed
  • Import → Discovery: When structure is clear but content is thin
  • Import → Drafting: When converting well-structured input
  • Drafting → Calibration: After each major section or full draft
  • Calibration → Drafting: After addressing flags

Blocking transitions:

  • Do NOT move Discovery → Drafting until convinces statement is filled
  • Do NOT draft slides without mapped threads
  • Do NOT finalize without calibration check

Multi-mode operation: Thread tracking runs alongside all modes. Mark threads even during drafting.

Semantic Tags Reference

Quick reference for <!-- AUDIO --> blocks. See references/semantic-tags-guide.md for full guidance.

| Tag | Purpose | When to Use | |-----|---------|-------------| | [HOOK] | Attention grabber | Opening of presentation | | [KEY_POINT] | Core insight | Main takeaway per slide | | [EVIDENCE] | Supporting data | Stats, research, proof | | [STORY] | Narrative example | Personal or case study | | [TRANSITION] | Flow connector | Between major sections | | [CALLBACK] | Reference earlier point | Reinforcing through-line | | [LANDING] | Closing message | Final impression | | [CTA] | Call to action | What to do next | | [PAUSE:ms] | Timing marker | After major points |

Integration Protocols

With prose-polish

When: After drafting narration blocks, before finalizing How: See integrations/prose-polish.md

  1. Invoke prose-polish on <!-- AUDIO --> content
  2. Flag scores >25 (presentation narration threshold)
  3. Cross-reference with presentation-blocklist
  4. Revise flagged passages for authenticity

With PresentationKit

When: After calibration, for generation

# Validate the talk track parses correctly
pk validate talk-track.md

# Generate outputs
pk build talk-track.md --output html,video,notes

Anti-Patterns

Don't:

  • Generate slides without interviewing first
  • Fabricate examples or data
  • Write generic narration then "customize" later
  • Use blocklist phrases in audio blocks
  • Create too many slides (less is more)
  • Forget the landing (audiences remember endings)

Do:

  • Interview before structuring
  • Use real examples from the speaker's experience
  • Apply semantic tags for clarity
  • Check timing (words vs. target_minutes)
  • Calibrate before finalizing
  • Test the hook—does it earn attention?

Example Session Flow

User: "I need to create a presentation about AI-native development for a tech conference"

You (Discovery): "Let's make sure this talk lands well. What should the audience do differently after your presentation? What action do you want them to take?"

User: [Responds about trying AI tools]

You (Thread): "Marking [CORE: Getting developers to actually try AI tools, not just hear about them]. What's the surprising insight—the one thing that changed your mind about AI-assisted development?"

User: [Shares insight about 70% time on non-coding]

You (Thread + Discovery): "Marking [EVIDENCE: 70% stat on developer time]. That's compelling. Is that from your experience or research you can cite?"

User: [Confirms Stack Overflow survey]

You (Transition check): "I can fill the gate now: 'This presentation convinces developers at TechConf to try AI tools this week by showing them how much time they waste on non-code tasks through the 70% stat and your team's 40% improvement.' Sound right?"

User: "Yes, let's draft"

You (Draft): "Starting the Talk Track v5. Based on our threads, I'm proposing 6 slides across 4 sections: Opening (hook), Problem (the 70% reality), Solution (AI-native approach), Closing (CTA). Here's the frontmatter and slides table..."

User: [Reviews]

You (Calibrate): "Running calibration... Format is valid. Timing: ~2100 words for 15 minutes—on target. One flag: slide 4 uses 'It's important to note' which is on the blocklist. Suggesting: 'Here's what matters:' or lead with the point directly."

State Management

Within session: Conversation history maintains state. Thread markers appear inline.

Across sessions: Export Talk Track v5 markdown. User can resume by providing the file.

For persistence: Consider using beads to track presentation development across sessions.

Escape Hatches

When things break, recover gracefully:

  • Thread explosion (10+): "We have too many threads for a clear talk. Let's pick the 3 that matter most."
  • Interview stalling: "I'm not getting the specificity I need. What's the one story you'd tell if you only had 5 minutes?"
  • Format confusion: User can say "show me the Talk Track so far" to see current state
  • Voice false positives: User can say "trust my voice on this" to skip calibration
  • Timing crisis: "This is running long. Cut slides 4-5 or compress to key points?"

Success Criteria

The true test: Can the user take this Talk Track v5 and confidently present it?

If yes → You're doing it right If no → Return to interview, check for generic content, verify evidence is real

References

  • references/talk-track-v5-spec.md - Complete format specification
  • references/semantic-tags-guide.md - Tag usage guidance
  • references/presentation-blocklist.md - AI patterns to avoid
  • references/format-detection.md - Input format recognition
  • integrations/prose-polish.md - Voice calibration protocol
  • assets/talk-track-v5-template.md - Starter template

Version: 1.0.0 Compatible with: Talk Track v5, v6 PresentationKit: See CLAUDE.md for CLI commands