Marp slides

MARP Slides

Startup

1. Read references/learnings.md — summarize the 3 most relevant bullets for this task
2. Detect mode from the request (see Mode Selector)

Mode Selector

Request signals	Mode
"create", "build", "make slides", "new deck", "presentation"	A: Create
"edit", "update", "change slide N", "revise", "fix"	B: Edit
"theme only", "CSS only", "just the styles", "color palette"	C: Theme
"convert from", "turn this into slides", "import", "use this outline"	D: Convert
"summarize this deck", "what does this cover", "read this marp"	E: Read

---

Mode A: Create

A1. Intake (ask all at once in a single message)

Present these 4 questions before generating anything:

1. Topic + purpose — what is the deck about, and who is the audience?
2. Slide count — how many slides? (if unsure, apply word-count heuristic from references/quality-rules.md)
3. Theme — dark (default) or light? Describe a vibe or brand if relevant.
4. Brand context — use existing brand colors/logo? (if yes, read your brand guidelines file)

Wait for reply before continuing.

Why: Generating before intent is locked is the #1 source of wasted cycles in deck authoring. The 4-question intake costs one round-trip and saves 3–5 regeneration loops.

A2. Theme + outline

1. Select theme using signal-to-theme mapping table in references/themes.md
2. Draft a slide-by-slide outline: slide title, one key idea, layout type, visual element planned
3. Present 3 layout direction options (e.g. minimal / data-heavy / editorial) — not full decks, just 3-slide skeleton previews as markdown
4. Wait for user to approve direction or request changes before generating the full deck

A3. Generate

1. Read references/components.md for SVG chart and component patterns
2. Read references/layout-patterns.md for rotation rules and composition guidelines
3. Read 2–3 examples from examples/ that most closely match the approved style
4. Generate the full .md MARP file

Hard rules during generation (enforce silently — do not annotate violations):

• Use - for list items, never * (asterisks trigger animation mode in Marp live preview)
• Rotate layout type every 2–3 slides — same template 3x in a row = quality failure
• Titles ≤35 characters; bullets ≤6 per slide; one key idea per slide
• Put context and detail in speaker notes (<!-- note: ... -->), not on slides
• Never add decorative accent lines directly under slide titles (telltale AI pattern)
• SVG backgrounds: always contain, never cover (cover crops charts at edges)

A4. QA self-review

1. Load references/quality-rules.md — run SlideGauge checklist against the generated deck
2. Fix any slide that fails silently
3. Report only if a slide required significant structural redesign

A5. Output + export

1. Save to presentations/[slug].md (or user-specified path)
2. Verify .marprc.yml exists at the project root — create from template in references/cli-guide.md if missing
3. Show the relevant export commands from references/cli-guide.md

---

Mode B: Edit

1. Read the existing .md file — parse slide boundaries (---)
2. Identify the target slide(s) from the request
3. Apply edits; re-check modified slides against references/quality-rules.md
4. Confirm layout rotation rule is not broken across surrounding slides

---

Mode C: Theme

1. Ask: brand colors? target mood? dark or light base?
2. Build CSS using 60-30-10 color rule and WCAG AA contrast — see references/themes.md
3. Present 3 palette directions before committing
4. Save theme CSS with /* @theme name */ header comment (required for CLI recognition)

---

Mode D: Convert

1. Parse input format (plain text, outline, bullet list, document)
2. Determine slide boundaries by topic/section shift, not by line breaks
3. Generate MARP .md applying Mode A generation rules
4. Run A4 QA self-review

---

Mode E: Read

1. Load the .md file
2. Parse slides, speaker notes, and frontmatter
3. Return: slide count, theme used, key idea per slide, layout types, SlideGauge quality flags

---

Closing Feedback Gate

After final output, ask once: "Did this produce what you needed? Any corrections or preferences I should remember?"

Route the response:

User says	Destination
Behavioral ("don't do X", "I prefer Y")	Append to references/learnings.md
Factual exception ("the flag is actually Z")	Append to references/edge-cases.md
"Never do X again"	Add rule to this SKILL.md
Explicit approval / "this was perfect"	Save output to examples/approved/
No response / "looks good"	Do nothing

---

<!-- Built with Agent Engineer Master — get your own production-ready skill: www.agentengineermaster.com/skill-engineer -->