Ppt Creator

ppt-creator

Produces a presentation-ready .pptx in the personal design system derived from RAG Done Right Production.pptx. Suitable for conference talks, meetups, internal presentations, and any non-work-branded decks.

All design values are sourced from references/design-system.md. Do not invent colour values, spacing, or typography outside that file.

Reference Files

Read only at the stage they are needed — never all at once.

| File | Read when | |------|-----------| | personal-info.md | At invocation — speaker name, title, QR path, defaults | | references/design-system.md | At invocation — C palette, primitives, layout grid | | references/slide-templates.md | At GENERATE only — layout skeletons | | references/boilerplate.js | At GENERATE — copy verbatim as deck file base, do not read into context | | references/infographic-workflow.md | Only when an infographic candidate is identified — do not load otherwise | | references/artifact-guide.md | At ARTIFACTS checkpoint only — export settings and embed rules for napkin.ai, draw.io, carbon.now.sh, Mermaid, NotebookLM | | references/design-system-override-guide.md | Only at DESIGN-DISCOVERY — when user requests a custom/alternative design system |

# Run at invocation:
cat ~/.claude/skills/ppt-creator/personal-info.md
cat ~/.claude/skills/ppt-creator/references/design-system.md

# Run only when GENERATE begins:
cat ~/.claude/skills/ppt-creator/references/slide-templates.md

# At GENERATE: copy the file, do not cat it
cp ~/.claude/skills/ppt-creator/references/boilerplate.js /path/to/[deck-name].js

# Run ONLY when an infographic candidate is identified (not before):
cat ~/.claude/skills/ppt-creator/references/infographic-workflow.md

# Run ONLY at the ARTIFACTS checkpoint (not before):
cat ~/.claude/skills/ppt-creator/references/artifact-guide.md

# Run ONLY at DESIGN-DISCOVERY (not before):
cat ~/.claude/skills/ppt-creator/references/design-system-override-guide.md

---

Workflow

RESEARCH → ALIGN [CHECKPOINT] → STRUCTURE (→ DECK-PLAN.md) → ARTIFACTS [CHECKPOINT] → DESIGN [CHECKPOINT] → GENERATE → REVIEW [CHECKPOINT] → CLOSE

---

Stage 1 — RESEARCH

First action: ask source mode. Do not search until answered.

RESEARCH SOURCE MODE
──────────────────────────────────────────────
How should I source content for this deck?

  A  Internet only
  B  Local files / documents you provide
  C  Hybrid — web first, you supplement

A / B / C ?
──────────────────────────────────────────────

Mode A: Run 5–8 web searches immediately.

Mode B: Say "Please share your files or paste your content." Wait. Do not search. Treat everything provided as primary source.

Mode C: Run 3–5 web searches. Present findings. Ask: "Anything to add, replace, or correct?" Incorporate user material as primary, overriding conflicts.

Research standard (A and C): Specific numbers, named sources, dateable claims. Not summaries. Cover: current state, quantitative benchmarks, failure modes, competitive context, audience priors.

Citation rule: Every web-sourced data point gets a superscript assigned here (¹ ² ³ …). Same number = same source throughout the entire deck. Internal / user-provided data = [Internal]. No citation number needed.

Synthesis output:

RESEARCH SYNTHESIS — [Topic]
──────────────────────────────────────────────────────────────────
SOURCE MODE: [A / B / C]

CORE FINDING:
One sentence — the single most important thing this research establishes.

KEY DATA POINTS:
¹  [Claim or metric] — [Why it matters]
   [Publication, Year] — [URL]

²  [Claim or metric] — [Why it matters]
   [Publication, Year] — [URL]

(5+ points for A/C; all available material for B)

NARRATIVE TENSION:
The gap between current state and what should be true.
This is what the deck resolves.

AUDIENCE LIKELY BELIEVES:
The assumption that needs to shift before the deck lands.

GAPS:
• What is unverified or missing
• What would strengthen the argument
──────────────────────────────────────────────────────────────────

After presenting, ask: "Does this capture what you need, or do you want to redirect, add material, or correct any findings?"

Wait. Re-number citations if data points change.

---

Checkpoint — ALIGN

Present all three questions with suggestions derived from the research. The user picks from the suggestions or writes their own — they should never have to start from a blank prompt.

Generate the suggestions from what the research actually revealed. Do not use generic placeholders — they must be specific to this topic and audience.

Before I build the structure, three quick questions. Pick one or write your own:

──────────────────────────────────────────────────────────────────
1. Who's the audience and what's the one thing they should walk
   away believing or deciding?

   A  [Specific audience + belief derived from research — e.g.
      "ML engineers at DevFest who should believe RAG outperforms
      fine-tuning for most production use cases"]
   B  [Alternative framing — different audience or belief]
   C  [A more action-oriented version — "should decide to..."]
   D  Write your own

──────────────────────────────────────────────────────────────────
2. What's one concrete thing they should leave able to DO?

   A  [Specific technique or action — e.g. "Implement a hybrid
      retrieval pipeline using BM25 + dense vectors"]
   B  [A decision or evaluation skill — e.g. "Evaluate whether
      their team should use RAG vs fine-tuning"]
   C  [Something smaller and more immediate]
   D  Write your own

──────────────────────────────────────────────────────────────────
3. What's the one thing you most want to avoid this deck implying?

   A  [A common misread of this topic — e.g. "That RAG is a
      drop-in replacement for a proper data pipeline"]
   B  [A credibility risk — e.g. "That this is only theoretical,
      not battle-tested"]
   C  [An audience alienation risk]
   D  Write your own
──────────────────────────────────────────────────────────────────

Wait for all three answers. Use Q1+Q2 to define INTENDED OUTCOME in the slide plan. Use Q3 to audit every slide for accidental implication before presenting the plan.

---

Stage 2 — STRUCTURE

Build the slide plan. This is where the narrative architecture happens.

A strong structure has:

• An opening that creates tension, not just announces a topic
• A credibility beat at slide 2 (ORANGE-SPEAKER) before any data is shown
• A data layer that earns the argument before making it
• A diagnosis slide that names what is actually broken, not just symptomatic
• Options or a path forward with trade-offs named honestly
• A close that ties back to the opening tension

Credibility beat — ORANGE-SPEAKER at slide 2: Include unless the user explicitly says to skip it or the event context makes it redundant (e.g., the speaker is already well-known to the entire room). Credentials must be specific and earned — never generic buzzwords.

Context detection — apply before choosing layouts: Detect audience context from the user's description and adjust layout emphasis:

• Developer/GDG/meetup → more DARK-NROW (techniques), code slides, architecture diagrams
• Enterprise/conclave/CXO → more LIGHT slides (frameworks), metrics, business impact framing
• Keynote/main stage → single bold claim per slide, max visual impact, minimal text
• Workshop/tutorial → step-by-step DARK-NROW heavy, code screenshots, higher slide count

Content density — enforce before presenting the plan: Every slide must stay within budget (from design-system.md). If a planned slide is over budget, split it or cut content before presenting to the user.

Infographic evaluation — run for every content slide:

After drafting the narrative structure, review each slide and ask: > "Would this slide be clearer and more memorable as a visual rather than text?"

Flag any slide that matches these patterns with [INFOGRAPHIC?] in the plan:

• 3+ data points forming a relationship, trend, or comparison
• A process flow with 4+ steps (especially with branching)
• A hierarchy or tree relationship
• Numeric-heavy content where >60% of the slide is numbers
• A before/after with quantified delta

For each flagged slide, include an infographic proposal block in the STRUCTURE output:

[INFOGRAPHIC?] Slide [N] — [current layout]
  Proposed: [chart type — why it is clearer than text]
  Tool:     [chartjs-node-canvas | Mermaid CLI]
  Data:     [the actual data points that would be visualised]
  Approve?  yes / no / describe change

Rules:

• Do NOT generate infographics speculatively — only after user approval
• Do NOT load infographic-workflow.md until at least one slide is approved
• If user declines all infographics, proceed with text layouts — no retry
• Infographic generation runs as an isolated sub-agent after STRUCTURE sign-off,

before GENERATE begins — see references/infographic-workflow.md

The slide plan is a suggestion. Every element is open to change: slide count, sequence, layouts, titles, content briefs. The only non-negotiable is the design system. Present it as a starting point.

If source mode was B (local only): Omit all Citations: lines from the plan. If source mode was A or C: Include citation refs per slide as shown.

SLIDE PLAN — [Topic]
──────────────────────────────────────────────────────────────────
NARRATIVE ARC:    [The one-sentence story this deck tells]
INTENDED OUTCOME: [What the audience does or believes after the last slide]

01  [LAYOUT]   [Role of this slide — not just its title]
    Content:   [What goes on it and why]
    Citations: [¹ Source, Year] or none

02  [LAYOUT]   [Role]
    Content:   [What and why]
    Citations: [refs] or none

... (continue for all slides)

──────────────────────────────────────────────────────────────────
Total: N   Orange: N   Dark: N   Light: N
Time at 2 min/slide: ~N minutes

CITATION INDEX:
¹  [Full source, Year] — [URL]
²  [Full source, Year] — [URL]
(all superscripts used — omit section if source mode B)

This is a suggestion. Tell me what to add, remove, reorder, or change
before I build. Slide count and structure are entirely yours to shape.
──────────────────────────────────────────────────────────────────

Wait for explicit confirmation or changes. Incorporate all edits.

After sign-off: write the approved plan to DECK-PLAN.md in the working directory. This file is what the user takes away to create external artifacts (napkin.ai infographics, draw.io diagrams, carbon.now.sh screenshots, etc.) before generation begins.

# After plan sign-off:
cat > /path/to/DECK-PLAN.md ---.pptx
# e.g.: rag-production-devfest-blr-indranil-2026.pptx
#       llm-evaluation-pycon-india-indranil-2026.pptx
#       data-mesh-conclave-indranil-2026.pptx

Rules:

• Derive TalkSlug from the talk title: 2–4 meaningful words, lowercase, hyphens
• Derive EventSlug from the event name: shorten to the distinctive part (e.g. "devfest-blr", "pycon-india", "conclave")
• AuthorSlug: first name only from personal-info.md (e.g. "indranil") — keep it short
• YYYY: year only — month is noise unless the same talk runs twice in a year
• If the title or event is short enough (≤ 4 words), use it verbatim; slug only when needed to keep the filename under 60 chars total

Now read slide-templates.md and copy boilerplate.js:

cat ~/.claude/skills/ppt-creator/references/slide-templates.md
cp ~/.claude/skills/ppt-creator/references/boilerplate.js /path/to/[deck-name].js

Code rules:

• One named function per slide: function s01_title() { ... }
• slideNum(s, N) — N is the actual slide number, not a placeholder
• Every addText() specifies fontFace: "Arial" — no exceptions
• Accent pill: call accent(s) first on every DS() and OS() slide
• Shadow factories: call mks() or mksl() inline — never assign to a variable and reuse
• Never reuse option objects across addShape/addText calls (pptxgenjs mutates)
• Every body text block with h > 0.30 must include lineSpacingMultiple: 1.15

— applies to: bullets, card body text, statement text, description blocks — skip for: slide numbers, metric numbers, short single-line labels (h ≤ 0.30)

• Speaker info: read from personal-info.md — expand ~ to absolute path for QR images

Speaker brand mark: After reading personal-info.md, call speakerMark(s, website) on every DS() and LS() body slide. Skip OS() slides. If website is blank in personal-info.md, omit the call entirely.

Code snippet slides: Use the DARK-CODE layout from slide-templates.md. The code block is a dark rectangle with monospace text runs. Each token type (keyword, string, comment, function name) gets its own colour run using pptxgenjs rich text arrays. Do NOT use carbon.now.sh unless the user explicitly requests a carbon screenshot — the DARK-CODE layout is native and editable in the .pptx.

Dark slide colour rules — enforced at code-write time:

• Body text: C.textOnDark (primary) or C.textBody (secondary) — NEVER C.textMid
• Captions/muted text: C.textMuted — NEVER C.textMid on dark slides
• No C.textOnDark or C.textBody on light slides — use C.textDark / C.textMid

Citation rules:

• Superscript Unicode inline on the data point: ¹ ² ³ ⁴ ⁵ ⁶ ⁷ ⁸ ⁹
• Call sourceFootnote(s, [...], onDark) on slides with web-sourced data
• Max 2 citations per sourceFootnote() call
• Full citation in speaker notes: SOURCE ¹: [Publication, Year] — [URL]
• Source mode B (internal only): no citations needed

Citations slide — mandatory for source mode A and C decks:

• ALWAYS generate a DARK-CITATIONS slide as the penultimate or final slide
• Use the sN_citations(allCitations) template from slide-templates.md

Speaker notes — every slide:

KEY STAT: [Most important number or claim on this slide]

TALKING POINTS:
• [What to say first — context]
• [Main point]
• [So-what / implication]
• [Anticipated objection]

SOURCES:                          ← omit if no web data on this slide
¹ [Publication, Year] — [URL]

Build and QA:

cd [working-dir]
npm install pptxgenjs --save 2>/dev/null || true
node [deck-name].js 2>&1

soffice --headless --convert-to pdf --outdir /tmp/ppt_qa [deck-name].pptx 2>/dev/null
pdftoppm -r 150 -png /tmp/ppt_qa/[deck-name].pdf /tmp/ppt_qa/slide

Build error handling — do not skip or paper over errors:

| Error | Cause | Fix | |-------|-------|-----| | Cannot find module 'pptxgenjs' | npm install failed or wrong dir | cd to correct dir, re-run npm install | | TypeError: pres.addSlide is not a function | pptxgenjs version mismatch | Check node_modules/pptxgenjs/package.json version — must be v3.x | | ENOENT: no such file or directory (image path) | QR or infographic path wrong | Expand ~ to absolute path, verify file exists | | RangeError / infinite loop in slide function | Mutated options object reused across calls | Find and fix the reused object literal | | soffice produces 0-page PDF | .pptx corrupt | Re-run node [deck].js — check for silent write errors | | pdftoppm produces no PNGs | soffice failed silently | Run without 2>/dev/null to see the error |

If any build error occurs: fix the root cause in the deck .js file and rebuild. Do not proceed to QA or REVIEW until node [deck].js exits with code 0.

QA checklist:

• [ ] Every title fits on one line, no wrapping
• [ ] No text or shape below y=5.09 (except slideNum and speakerMark)
• [ ] No element overlap within any slide
• [ ] accent() called first on every DS() and OS() slide
• [ ] Slide numbers sequential at x=9.25 y=5.09
• [ ] No blank slides
• [ ] No C.textMid on any dark/orange slide (grep the code)
• [ ] Citations slide present as last slide (source mode A/C only)
• [ ] All body text blocks with h > 0.30 have lineSpacingMultiple: 1.15
• [ ] Shadow factories called inline (never reused)
• [

…

Source & license

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

• Author: indranildchandra
• Source: indranildchandra/claude-local-starter
• License: MIT

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