Slide maker

<Purpose> End-to-end slide creation: from raw sources or a topic to a polished presentation. Handles research, synthesis, outlining, and slide generation. Produces self-contained reveal.js HTML by default (easy to edit, preview in any browser). Can delegate to pptx skill when PowerPoint output is requested. </Purpose>

<Use_When>

• User asks to create slides or a presentation about a topic
• User provides sources (URLs, PDFs, repos, vault notes) and wants slides
• User wants to edit an existing HTML or pptx slide deck
• User uses /slides or /slide-maker </Use_When>

<Do_Not_Use_When>

• User wants a research note, not slides (use /research)
• User just wants to read/extract text from a pptx (use pptx skill directly) </Do_Not_Use_When>

<Workspace> All intermediate files go in `temp/slides/` (gitignored). Structure: ``` temp/slides/ ├── research/ # Source research files (one per source) ├── images/ # Downloaded images (diagrams, charts, screenshots) ├── outline.md # Approved outline └── output.html # Draft before copying to notes/slides/ ```

Create workspace + output directories at the start of every session:

mkdir -p temp/slides/research temp/slides/images notes/slides

Final output always goes to notes/slides/<presentation-name>.html (or .pptx). Images used by the presentation are copied to notes/slides/images/. The temp/slides/ workspace is for intermediate files only. </Workspace>

<Steps>

Stage 1: PARSE — Classify Inputs

Parse $ARGUMENTS and any user context. Classify each input:

Input type	Detection	Strategy
URL	starts with http(s)://	WebFetch + download images
Git repo	github.com or git URL	clone to temp, read key files
PDF	.pdf extension	Read tool (for local) or download
Vault note	(Type) Name or [[wikilink]]	Read via MCP or Read tool
Existing slides (.html)	.html extension	Editing mode
Existing slides (.pptx)	.pptx extension	Delegate to pptx skill
Raw topic	none of the above	WebSearch + WebFetch

Auto-detect research depth:

• Quick mode: 1-2 simple sources or a short topic → inline research, no subagents
• Thorough mode: 3+ sources, complex topic, or broad request → parallel research agents

Present what was detected:

Detected sources:
1. [URL] blog post — will fetch + extract
2. [URL] paper PDF — will download and read
3. "transformer architecture" — will web search

Mode: thorough (3 sources → parallel research)
Proceeding to research...

Don't wait for confirmation unless something is ambiguous.

Stage 2: DISCOVER — Understand the Goal

Before researching, understand what the user actually needs. Ask up to 5 focused questions in a single message. Adapt questions to what's already known — skip anything the user already specified.

Questions to consider (pick the ones that aren't already answered):

1. Goal & context — What's this presentation for? (class lecture, team meeting, conference talk, pitch, self-study, sharing with a friend?)
2. Audience — Who's watching? (technical experts, general audience, exec stakeholders, students?) What do they already know?
3. Language — What language should the slides be in? (Don't assume English — the user may want slides in their native language or a mix)
4. Angle & emphasis — What's the ONE thing the audience should walk away with? Any specific aspects to emphasize or skip?
5. Vibe & format — Formal or casual? Dense/technical or visual/storytelling? Any color or style preferences? Dark or light theme?

How to ask — be conversational, not interrogation-style:

Before I start researching, a few quick questions so I build the right deck:

1. What's this for — conference talk, team meeting, class?
2. Who's the audience? What do they already know about [topic]?
3. What language should the slides be in?
4. Any specific angle — what should the audience walk away remembering?
5. Style preference — formal/minimal, bold/visual, dark/light theme?

(Skip any that don't apply — I'll use sensible defaults)

After the user responds, synthesize their answers into a creative brief that guides all downstream stages. Save it to temp/slides/brief.md:

# Creative Brief

**Topic**: [topic]
**Goal**: [what this presentation accomplishes]
**Audience**: [who, what they know, what they care about]
**Language**: [language for slide content]
**Key takeaway**: [the ONE thing the audience should remember]
**Tone**: [formal / casual / technical / storytelling / provocative]
**Style direction**: [dark/light, color mood, density level]
**Emphasis**: [specific aspects to highlight]
**Skip**: [anything the user explicitly wants to avoid]
**Format**: [reveal.js HTML / pptx]
**Duration**: [if mentioned — affects slide count]

Palette selection: Based on the brief, choose a palette from the Design Principles table (or craft a custom one). The palette should feel designed for THIS topic and THIS audience — not generic. Include your palette choice and reasoning in the brief.

Skip this stage when:

• Editing an existing deck (the design decisions are already made)
• The user explicitly says "just do it" or provides all context upfront

Then proceed to research with the brief guiding what to look for.

Stage 3: RESEARCH — Gather Information

Thorough mode (3+ sources or complex topic)

Read the agent definition:

Read("${CLAUDE_SKILL_DIR}/agents/source-researcher.md")

Spawn one source-researcher agent per source, all in parallel in a single message. Each agent:

• Receives: one source, its output file path, images directory
• Writes structured research to temp/slides/research/source-NN-slug.md
• Downloads relevant images to temp/slides/images/
• Returns a 3-5 sentence summary (keeps orchestrator context small)
• Uses sonnet model

Agent(
  model="sonnet",
  prompt="You are Source Researcher. Follow these instructions:

  [INSERT FULL CONTENT OF agents/source-researcher.md]

  SOURCE: [url/path/topic]
  SOURCE_NUMBER: [NN]
  OUTPUT_FILE: temp/slides/research/source-NN-slug.md
  IMAGES_DIR: temp/slides/images/

  Research this source and write findings to the output file.
  Return only a 3-5 sentence summary."
)

If a source fails, continue with what succeeded — note the failure.

For raw topics with no other sources, spawn one agent that does WebSearch (2-3 queries) + WebFetch (top 3 results).

Quick mode (1-2 simple sources)

Skip agents. Do inline research:

1. WebFetch each URL (or Read each file)
2. For raw topics: WebSearch + WebFetch top 2-3 results
3. Download key images to temp/slides/images/
4. Write a single research file to temp/slides/research/source-01-topic.md

Image handling

During research (both modes), download diagrams, charts, and screenshots that could enhance slides. Skip decorative/stock images. Save to temp/slides/images/ with descriptive names.

curl -sL "IMAGE_URL" -o temp/slides/images/descriptive-name.png

Stage 4: OUTLINE — Synthesize + User Approval (MANDATORY)

Read all research files from disk:

Glob(pattern="temp/slides/research/*.md")
# Then Read each file

Read the creative brief from temp/slides/brief.md — use it to guide content selection, ordering, emphasis, language, and tone.

Build a slide outline. Present it to the user:

## Slide Outline

**Title**: [Presentation Title — in the brief's language]
**Audience**: [from brief]
**Language**: [from brief]
**Tone**: [from brief]
**Palette**: [from brief — name + hex values]
**Key takeaway**: [the ONE thing from the brief]
**Slides**: ~[N] slides

---

1. **Title Slide** — [title + subtitle]
2. **[Insight headline, not label]** — [2-3 key points] | Layout: [type]
3. **[Insight headline]** — [key points] | Image: [image-name.png]
4. ...
N. **Closing / Key Takeaways** — [summary points]

Headline rule: every slide headline should be an insight ("Attention is $O(n^2)$ — that's the bottleneck"), not a label ("Attention Mechanism").

ALWAYS wait for user approval before proceeding to Stage 4.

The user may:

• Approve as-is → proceed
• Request changes → revise and re-present
• Add/remove slides → adjust

Save the approved outline to temp/slides/outline.md.

Stage 5: CREATE or EDIT

Path A: New reveal.js HTML deck (default)

Read the agent definition and template:

Read("${CLAUDE_SKILL_DIR}/agents/slide-writer.md")
Read("${CLAUDE_SKILL_DIR}/references/reveal-template.html")

Launch the slide-writer agent:

Agent(
  model="sonnet",
  prompt="You are Slide Writer. Follow these instructions:

  [INSERT FULL CONTENT OF agents/slide-writer.md]

  CREATIVE BRIEF:
  [content of temp/slides/brief.md]

  OUTLINE:
  [approved outline from temp/slides/outline.md]

  RESEARCH FILES (read each for detailed content):
  [list of temp/slides/research/*.md paths]

  IMAGE FILES AVAILABLE:
  [list of temp/slides/images/* paths]

  TEMPLATE:
  [reveal-template.html content]

  OUTPUT_FILE: temp/slides/output.html

  Build the presentation following the brief's language, tone, palette,
  and audience direction. Write the HTML to the output file."
)

Path B: New pptx deck (when user explicitly requests PowerPoint)

Option 1 — Convert from HTML (preferred, faster): First create the reveal.js HTML (Path A), then convert to pptx:

npx --yes -p pptxgenjs node "${CLAUDE_SKILL_DIR}/scripts/html2pptx.js" notes/slides/name.html notes/slides/name.pptx

This preserves palette, layouts, and content from the HTML. Review the output and fix any issues.

Option 2 — Direct pptx creation (for complex cases): Delegate to the pptx skill at .agents/skills/pptx/. Pass:

• Approved outline
• Research file paths
• Image paths
• Design choices from outline (palette, tone)

Path C: Edit existing HTML slides

1. Read the existing HTML file
2. Apply requested changes using the Edit tool (HTML is plain text — trivial)
3. For structural changes (add/remove slides): read, modify section, write back

Path D: Edit existing pptx

Delegate to pptx skill's unpack/edit/pack workflow.

Stage 6: PREVIEW + ITERATE

After the slide-writer agent finishes:

1. Copy the HTML + images to their final location:

# Generate a slug from the presentation title
SLUG="presentation-name"  # kebab-case from title
cp temp/slides/output.html "notes/slides/${SLUG}.html"
# Copy any images referenced by the presentation
cp -r temp/slides/images/ "notes/slides/images/" 2>/dev/null || true

2. Tell the user:

Presentation created: notes/slides/[name].html

Open in your browser to preview (File → Open or drag into browser).
Let me know if you'd like any changes — I can edit individual slides,
adjust colors, reorder, add/remove content.

Need PowerPoint? I can export to .pptx too.

For iteration:

• HTML edits use the Edit tool directly on notes/slides/[name].html
• Fast — no rebuild needed, HTML is plain text
• Track which slides the user wants changed
• Apply targeted edits, not full regeneration

</Steps>

<Design_Principles> Borrowed from the pptx skill — apply to ALL slide formats.

Color Palettes

Pick a palette that matches the topic. Don't default to generic blue.

Theme	Primary	Secondary	Accent
Midnight Executive	#1E2761 (navy)	#CADCFC (ice blue)	#FFFFFF (white)
Forest & Moss	#2C5F2D (forest)	#97BC62 (moss)	#F5F5F5 (cream)
Coral Energy	#F96167 (coral)	#F9E795 (gold)	#2F3C7E (navy)
Warm Terracotta	#B85042 (terracotta)	#E7E8D1 (sand)	#A7BEAE (sage)
Ocean Gradient	#065A82 (deep blue)	#1C7293 (teal)	#21295C (midnight)
Charcoal Minimal	#36454F (charcoal)	#F2F2F2 (off-white)	#212121 (black)
Teal Trust	#028090 (teal)	#00A896 (seafoam)	#02C39A (mint)
Berry & Cream	#6D2E46 (berry)	#A26769 (dusty rose)	#ECE2D0 (cream)
Cherry Bold	#990011 (cherry)	#FCF6F5 (off-white)	#2F3C7E (navy)

Layout Rules

• Every slide needs a visual element — image, chart, diagram, large number, or shape
• Vary layouts — don't repeat the same structure across slides
• Dominance over equality — one color dominates (60-70%), others support
• Dark/light contrast — dark backgrounds for title + closing, light for content
• Layout options: two-column, icon+text rows, grids, half-bleed images, stat callouts, timelines

Typography (for reveal.js)

• Use Google Fonts for variety (loaded via CDN in template)
• Title slides: 2.5-3em bold
• Slide headlines: 1.5-2em
• Body text: 0.9-1.1em
• Size contrast matters — headlines must visually dominate body text

Avoid (Common Mistakes)

• Don't repeat the same layout across slides
• Don't center body text — left-align; center only titles
• Don't use low-contrast text
• Don't create text-only slides — add visual elements
• Don't default to blue — match the topic
• NEVER use accent lines under titles — hallmark of AI slides
• Don't cram too much on one slide — fewer points, more slides </Design_Principles>

<Tool_Usage>

• WebSearch: Find sources for raw topics
• WebFetch: Fetch URL content, extract insights
• Read/Glob: Read local files, vault notes, research files
• Agent: Parallel source-researcher agents (sonnet), slide-writer agent (sonnet)
• Write: Create output HTML file, research files
• Edit: Iterate on existing HTML slides
• Bash: mkdir, curl (image downloads), file operations </Tool_Usage>

<Examples> <Good> User: /slides about transformer architecture 1. Parse → raw topic, quick mode 2. Discover → asks: who's the audience? language? what angle? → user says "ML study group, English, focus on why attention replaced RNNs, dark theme" 3. Brief → audience: ML practitioners, key takeaway: attention enables parallelism, palette: Ocean Gradient (matches deep learning vibe), tone: technical but visual 4. Research → WebSearch 3 queries, fetch top results, download architecture diagram 5. Outline → 10 slides ordered around the "why attention wins" narrative 6. User approves with minor reorder 7. Create → reveal.js HTML, dark Ocean Gradient, diagrams embedded, speaker notes 8. Preview → user asks to add flash attention slide → Edit tool, done

User: ทำสไลด์จาก [url1] [url2] [url3] กับ (Paper) Scaling Laws

1. Parse → 3 URLs + 1 vault note, thorough mode
2. Discover → asks: ภาษาอะไร? กลุ่มเป้าหมาย? → user says "Thai slides, for team meeting, casual tone, bright colors"
3. Brief → language: Thai, audience: engineering team, palette: Coral Energy, tone: casual/visual
4. Research → 4 parallel agents, each writes research file
5. Outline → 12 Thai-language slides synthesizing across all sources
6. User approves
7. Create → slide-writer produces Thai HTML with Coral Energy palette </Good>

<Bad> - Skipping DISCOVER and assuming English, formal, blue theme - Skipping the outline checkpoint and going straight to slides - Putting all research content in agent prompts instead of writing to disk - Creating text-only slides with no visual elements - Using the same two-column layout for every slide - Not downloading available images during research phase - Ignoring user's language preference in the final slides </Bad> </Examples>

<Escalation_And_Stop_Conditions>

• Too many sources (>8): Ask user to prioritize top 5-6
• Source fetch fails: Note failure, continue with remaining sources
• User wants pptx: Delegate to pptx skill after outline approval
• Ambiguous topic: Ask one clarifying question about audience/angle
• Existing deck format unclear: Ask user to confirm format </Escalation_And_Stop_Conditions>

$ARGUMENTS