Shuttleslide

Bidirectional PowerPoint ↔ HTML conversion with round-trip format preservation — plus an optional AI slide-generation pipeline.

Shuttleslide is a Python library and CLI that turns .pptx files into semantic, navigable HTML and back. Unlike one-way converters that flatten slides into images or fixed templates, Shuttleslide emits real, editable elements on both sides of the round trip — text boxes, tables, native DrawingML shapes, gradients — and threads data-pptx-* metadata through every conversion so your formatting survives the return journey.

---

Table of contents

• What is Shuttleslide?
• Why Shuttleslide?
• How it compares
• Installation
• Quick start
• Examples
• Python API
• CLI reference
• Architecture
• Design philosophy: CLI-first, skill-ready
• Roadmap
• Contributing
• Acknowledgments
• License

---

What is Shuttleslide?

Three capabilities, one package, each independently usable:

1. PPTX → HTML (slidecraft to-html)

Convert any .pptx into a single self-contained HTML file. Choose between three layout modes:

• flow — semantic, scrollable page. Best for RAG / web publishing / SEO.
• pptview — PowerPoint-style editor layout. Best for design fidelity.
• slideshow (default) — interactive presentation with keyboard navigation, slide transitions, and entrance animations.

Text formatting, tables, images, shapes, gradients, bullets, and master-theme styling are all preserved.

2. HTML → PPTX (slidecraft to-pptx)

Convert HTML back into a .pptx with native DrawingML shapes — every element remains directly clickable and editable in PowerPoint. No LLM required: a deterministic rule-based pipeline (Playwright layout extraction → PPT-DSL JSON → python-pptx) handles the work.

The HTML subset Shuttleslide accepts on this side is intentionally limited to a "Slide HTML" vocabulary — see docs/round-trip.md for what's supported.

3. AI slide generation (slidecraft generate, slidecraft review)

Skip the source file entirely — give the agent a topic and it produces a multi-slide deck via an OpenAI-compatible LLM (Zhipu GLM, DeepSeek, OpenAI, vLLM, Ollama, …). review opens a web UI where each pipeline stage pauses for human approval.

Requires the [ai] (and [review] for the web UI) install extras. See Installation.

---

Why Shuttleslide?

• Round-trip preservation is the goal, not a side effect. data-pptx-* attributes are emitted on PPTX → HTML and consumed on HTML → PPTX, so a deck that goes out and comes back keeps its coordinates, fonts, and structure. This is the differentiator — no other open-source tool does both directions.
• Native DrawingML, not flattened images. HTML → PPTX produces real PowerPoint shapes (text boxes, paths, gradients), built on a vendored copy of ppt-master's svg_to_pptx engine. The result remains natively editable in PowerPoint.
• Deterministic when you want it, AI when you don't. The to-html / to-pptx pipeline is rule-based — no network, no LLM, no flakiness. generate is the optional AI path for when you don't have a source deck at all.
• PPT-aware typography. Line-height, paragraph spacing, and font-shrink-on-overflow are calibrated against real PowerPoint rendering (see docs/round-trip.md for the empirical ratios). HTML output looks like the original, not like a CSS approximation of the original.
• Three layout modes cover three different use cases — RAG ingestion, design-faithful preview, and live presentation — from one parser.
• Extension points. Four entry-point groups let external packages add CLI commands, pipeline stages, or substitute prompts without touching this repo.

---

How it compares

Capability	Shuttleslide	ppt-master	pptx-to-html5	python-pptx	Aspose.Slides	LibreOffice
PPTX → HTML	✅ 3 layouts	❌	✅	❌	✅	✅
HTML → PPTX	✅	✅ (AI-driven)	❌	❌	❌	❌
Round-trip metadata	✅	❌	❌	❌	❌	❌
Native DrawingML output	✅	✅	❌	⚠️ manual	✅	✅
Works without LLM	✅	❌	✅	✅	✅	✅
Open source & free	✅ MIT	✅ MIT	✅	✅ MIT	❌ commercial	✅
Install via pip	✅	❌ git clone	✅	✅	✅	❌ system pkg

Shuttleslide vs ppt-master — the two projects are complementary, not competing. ppt-master is an AI workflow skill that runs inside Claude Code / Cursor / Copilot CLI: you chat with the agent and it generates a deck from source documents. Shuttleslide is a Python library + CLI you call from code or shell. Shuttleslide vendors ppt-master's SVG → DrawingML converter under src/shuttleslide/_vendored/svg_to_pptx/ (MIT-licensed, with attribution) — see ACKNOWLEDGMENTS.

A detailed breakdown of trade-offs is in docs/comparison.md.

---

Installation

# Core: PPTX ↔ HTML conversion (to-html, to-pptx, analyze)
pip install shuttleslide

# Add AI slide generation (generate command)
pip install shuttleslide[ai]

# Add the web review UI (review command)
pip install shuttleslide[review]

# Everything
pip install shuttleslide[all]

One-time setup for Playwright's headless Chromium (used by HTML → PPTX layout extraction and PPTX → HTML text-fit measurement):

python -m playwright install chromium

Python 3.9+ is required. Python 3.8 has reached end-of-life and is not supported.

From source (development)

git clone https://github.com/solider-shuwen/shuttleslide.git
cd shuttleslide
pip install -e ".[dev]"
python -m playwright install chromium

---

Quick start

A. PPTX → HTML

# Interactive slideshow (default)
slidecraft to-html deck.pptx -o presentation.html

# Scrollable single page
slidecraft to-html deck.pptx -o presentation.html --layout flow

# Editor-style preview
slidecraft to-html deck.pptx -o presentation.html --layout pptview

B. HTML → PPTX

to-pptx is a deterministic rule-based pipeline — no LLM, runs locally. It works best on HTML that conforms to Shuttleslide's "Slide HTML" subset, which is exactly what generate produces. The recommended round-trip path is generate → to-pptx:

# 1. Generate a deck whose HTML matches the Slide HTML subset
slidecraft generate "Quarterly business review" \
    --style business -o tmp/qbr/

# 2. Convert each generated slide to a native PPTX
slidecraft to-pptx tmp/qbr/1.html -o tmp/qbr/1.pptx

The resulting .pptx is fully editable in PowerPoint — every text box, shape, gradient, and SVG-converted path is a native DrawingML element, not a flat image.

You can also feed hand-written or third-party HTML to to-pptx, but fidelity depends on how closely the markup follows the Slide HTML subset. Arbitrary web pages are explicitly out of scope (see CLAUDE.md — iron rules).

# Arbitrary "Slide HTML" subset HTML → PPTX (best-effort)
slidecraft to-pptx slides.html -o output.pptx

C. AI generation (optional)

slidecraft generate "Introduction to Machine Learning" \
    --api-base https://api.deepseek.com \
    --api-key your-key-here \
    --vlm-api-base https://open.bigmodel.cn/api/paas/v4 \
    --vlm-api-key your-key-here \
    --model deepseek-v4-flash \
    --vlm-model glm-4.6v \
    -o tmp/gen/

Or launch the web review UI for human-in-the-loop generation:

pip install shuttleslide[review]
slidecraft review

D. Use with Claude Code (and other AI agents)

Shuttleslide ships an official Claude Code skill bundle inside the pip package. One command deploys three skills (slidecraft, slidecraft-to-html, slidecraft-review) to ~/.claude/skills/, after which Claude Code — or any agent harness that consumes the SKILL.md format — can invoke slidecraft via natural language:

pip install shuttleslide
slidecraft install-skill          # deploys to ~/.claude/skills/

Then in Claude Code:

"Convert this deck to HTML for RAG ingestion" → triggers slidecraft-to-html "Review the generated slides before finalizing" → triggers slidecraft-review

Idempotent (sha256-checked); re-run after upgrading Shuttleslide to pick up skill updates, or pass --force to overwrite local edits.

---

Examples

The example/ directory contains reproducible end-to-end samples. Each subdirectory has its own README describing the input, the exact command used, and where the output lands.

example/ai-generated/ — "How Shuttleslide Works"

A 6-slide deck about Shuttleslide itself, generated end-to-end by slidecraft generate against an OpenAI-compatible LLM.

slidecraft generate "How Shuttleslide Works: The Round-Trip Story" \
    --style tech --slides 6 -o example/ai-generated/

example/pptx-to-html/ — PPTX → HTML

A real-world PPTX (with bullets, multi-level paragraphs, and SVG/shape assets) converted to an interactive slideshow HTML.

slidecraft to-html example/pptx-to-html/1.pptx

example/html-to-pptx/ — HTML → PPTX

A hand-written Slide-HTML converted to a native, editable PPTX.

slidecraft to-pptx example/html-to-pptx/sample.html

---

Web Review UI — human-in-the-loop generation

slidecraft generate runs the full pipeline end-to-end without intervention. slidecraft review runs the same pipeline but pauses after every stage so you can inspect the snapshot, request changes, and only approve when you're happy. The agent won't proceed until you say go — or until you edit the snapshot yourself.

This is the workflow to reach for when the deck matters: brand reviews, stakeholder approvals, decks that have to land on someone's calendar next Monday. You stay in the loop on every stage of theme design, outline, slide content, and rendering.

Starting a run

pip install shuttleslide[review]
slidecraft review

The config screen lets you set topic, style, target slide count, and (if not locked via .env or CLI flags) LLM credentials. Mock mode (--mock) runs the whole UI against canned state — useful for demos and screenshots without an LLM endpoint.

Stage-by-stage review

Once you click Start, the pipeline runs stage by stage. Each stage emits a snapshot you can review:

• Theme designer — color palette, gradients, typography pairing
• Outline planner — slide-by-slide structure and headlines
• Slide builder — element-level content per slide (text, SVG, image, table)
• Renderer — final HTML output

For every stage you have three options:

Action	What happens
Approve	The pipeline advances to the next stage using the current snapshot as-is.
Edit manually	Open the snapshot in an in-browser editor and change values directly — tweak copy, swap colors, fix an outline item. Your edits become the new truth the next stage reads from.
Ask the LLM	Describe what you want changed in natural language ("drop the third slide", "make the title punchier", "use a cooler palette") and the agent regenerates the affected parts of the snapshot. The LLM only sees the current snapshot plus your instruction, not the full conversation.

Both edit paths produce a new snapshot that you re-review — so you can iterate manually for some changes and let the LLM handle others, in any order, on the same stage.

Resuming

Review runs are persisted under tmp/web_review/<timestamp>/. If you close the browser or kill the server mid-run, restarting slidecraft review offers to resume from the last approved stage. Nothing is lost.

---

Python API

All three capabilities are also exposed as importable Python APIs.

PPTX → HTML

from shuttleslide.pptx_to_html import PPTXParser, FlowLayout

parser = PPTXParser("deck.pptx")
slides = parser.parse()
html = FlowLayout().convert(slides)

Path("out.html").write_text(html, encoding="utf-8")

HTML → PPTX

import asyncio
from shuttleslide.html_to_pptx import RuleSlideTransformer, PPTXRenderer

async def convert(html_path: str, pptx_path: str):
    html = Path(html_path).read_text(encoding="utf-8")
    transformer = RuleSlideTransformer()
    dsl = await transformer.transform_html(html, base_dir=Path(html_path).parent)
    PPTXRenderer(base_dir=Path(html_path).parent).render(dsl, pptx_path)

asyncio.run(convert("slides.html", "out.pptx"))

AI generation

from shuttleslide.agent import generate_slides

await generate_slides(
    topic="Introduction to Machine Learning",
    style_hint="business",
    output_dir="tmp/gen/",
)

Full API reference: docs/python-api.md.

---

CLI reference

Command	Purpose
slidecraft to-html	PPTX → HTML (flow / pptview / slideshow)
slidecraft to-pptx	HTML → PPTX (deterministic, no LLM)
slidecraft json-to-pptx	PPT-DSL JSON → PPTX directly
slidecraft analyze	Inspect PPTX structure
slidecraft generate	AI: topic → multi-slide HTML deck
slidecraft review	AI: launch the human-in-the-loop web UI
slidecraft warm-cache	Pre-download CDN assets for offline generation
slidecraft install-skill	Deploy bundled Claude Code skills to ~/.claude/skills/
slidecraft info	Print version and project info

Complete options, environment variables, and examples: docs/cli-reference.md.

---

Architecture

Round-trip pipeline (PPTX ↔ HTML)

flowchart LR
    P[(.pptx)] --> Parser[PPTXParser]
    Parser --> Slides[ParsedSlide array]
    Slides --> Layout{{Layout engine}}
    subgraph LayoutModes[Modes]
        L1[flow]
        L2[pptview]
        L3[slideshow]
    end
    Layout --> LayoutModes
    LayoutModes --> HTML["HTML + data-pptx-* metadata"]
    HTML --> Rule["RuleSlideTransformer<br/>(Playwright extraction + rules)"]
    Rule --> DSL[PPT-DSL JSON]
    DSL --> Renderer[PPTXRenderer]
    Renderer --> P2[(.pptx<br/>native DrawingML)]

PPTX → HTML writes data-pptx-* attributes on every converted element. HTML → PPTX can read those attributes to restore the original layout precisely. The whole loop is bidirectional and lossless at the semantic level.

AI generation pipeline (topic → slides)

flowchart LR
    Topic[Topic + style hint] --> Theme[Theme designer<br/>1 LLM call]
    Theme --> Outline[Outline planner<br/>1 LLM call]
    Outline --> Builder[Slide builder<br/>N LLM calls, 1 per slide]
    Builder --> Renderer[Jinja2 renderer<br/>deterministic, no LLM]
    Renderer --> Out[slide HTML<br/>Slide HTML subset]
    Out -. optional .- RT[RuleSlideTransformer]
    RT -.-> PPTX[(.pptx)]

slidecraft generate runs the four stages end-to-end. slidecraft review runs the same stages but pauses after each one for human review (see Web Review UI).

More detail: docs/round-trip.md and docs/comparison.md.

---

Design philosophy: CLI-first, skill-ready

Every Shuttleslide capability is a standalone slidecraft subcommand. This isn't an accident.

The CLI is the contract. Each command has a single, well-defined job — to-html converts a file, analyze inspects one, generate runs the AI agent — and produces deterministic outputs you can pipe, chain, or script. That means Shuttleslide isn't just a Python library: it's an AI-agent-ready toolkit. Any agent harness (Claude Code, Cursor, Copilot CLI, Cline, Aider, …) can read slidecraft --help and start orchestrating presentation workflows without learning a Python API.

This isn't theoretical. slidecraft install-skill deploys a bundled Claude Code skill package — three SKILL.md files that teach Claude Code the natural-language triggers for to-html, review, and the rest of the CLI. Other agent harnesses that consume the same format can use the same files. See Quick start §D above.

---

Roadmap

• Phase 1 — PPTX → HTML ✅ shipped
  • Text, tables, images, shapes, gradients, master-theme styling
  • Three layout modes (flow / pptview / slideshow)
  • CLI, Python API, deterministic (no LLM)
• Phase 2 — HTML → PPTX ✅ shipped
  • Slide-HTML subset, rule-based extraction, vendored SVG → DrawingML
  • Native, editable PowerPoint output
• Phase 3 — Round-trip closure 🚧 in progress
  • data-pptx-* metadata fully round-trips for all element types
  • Format-integrity verification tooling
• Phase 4 — Skill package 🚧 in progress
  • ✅ slidecraft install-skill ships three Claude Code skills inside the pip package
  • 📋 Marketplace manifest for one-click IDE installation

---

Contributing

Contributions are welcome. To set up a dev environment:

git clone https://github.com/solider-shuwen/shuttleslide.git
cd shuttleslide
pip install -e ".[dev]"
python -m playwright install chromium
pytest

Please open an issue first for significant changes, so we can align on scope before you spend time on a PR.

---

Acknowledgments

Shuttleslide builds on outstanding prior work:

• ppt-master by Hugo He — the SVG → native DrawingML conversion engine (src/shuttleslide/_vendored/svg_to_pptx/) is a vendored, MIT-licensed copy of his work. Without it, the HTML → PPTX direction wouldn't be possible at this quality.
• python-pptx — the foundation for reading and writing .pptx files.
• Click — CLI framework.
• Playwright — headless browser automation for layout extraction and text measurement.

---

License

MIT © Shuttleslide Contributors