Clone Any .pptx Template Into a New Deck With gpt-image-2

Template decks die in the fill-in-the-text step: you inherit someone's layout and then spend an hour fighting text boxes. gpt-image2-ppt-skills takes the opposite route — instead of filling text into traditional templates, it uses the visual taste, composition, and layout strengths of OpenAI's gpt-image-2 to generate each slide as a complete visual composition, then hands you a packaged .pptx.

Why This Skill Matters

The skill works natively in Claude Code, Codex, OpenClaw, Hermes, and other Skill-compatible agents (the quick start also names Cursor and Trae). One natural-language prompt yields 16:9 high-res images per slide plus a ready-to-send .pptx; uploaded .pptx files become style sources — the agent follows the template's layout, palette, and illustration language, then swaps in your content. Two details set it apart. It ships a template system: 265 templates as of the 2026-08-02 changelog entry, browsable in the SlideCraft Template Gallery with copy-ready prompts. And its lineage is documented in the acknowledgements: it started as a rework of op7418/NanoBanana-PPT-Skills, replacing Nano Banana Pro with OpenAI's gpt-image-2.

Editing is where image-first decks usually fall over, and this is the project's most engineered part. The default mode keeps each slide a full-bleed image. An opt-in editable mode — off unless you explicitly ask — reconstructs text, basic shapes, connectors, and hero artwork as separate PowerPoint objects. The shipped case study renders back at 13 selectable objects: 5 native text boxes, 6 native shapes, 1 clean plate, and 1 movable, resizable mascot-and-ice-cream image layer.

Installation

The README's preferred path is a sentence, not a command. Send this to your agent:

Please install gpt-image2-ppt-skills: https://github.com/JuneYaooo/gpt-image2-ppt-skills

For manual installs, the README documents:

git clone git@github.com:JuneYaooo/gpt-image2-ppt-skills.git
cd gpt-image2-ppt-skills
bash install_as_skill.sh --target claude   # Claude Code
# 或
bash install_as_skill.sh --target codex    # Codex

The script installs into ~/.claude/skills/gpt-image2-ppt-skills/ or ~/.codex/skills/gpt-image2-ppt-skills/. Direct API mode needs OPENAI_API_KEY (plus optional OPENAI_BASE_URL and quality settings); in Codex, if the agent has native image generation, the README's SKILL.md native path needs no key. One software requirement applies to the advanced modes: template cloning and editable mode need an executable local PPTX renderer — Windows PowerPoint, macOS Keynote, or LibreOffice. The skill checks this itself and explains how to install a compatible renderer when one is missing.

Real Workflow: Reskin a Company Template

1. Upload company-template.pptx and describe the deck:

I uploaded company-template.pptx. Follow its layout, palette, and visual language to make an 8-slide deck about our annual product strategy.

2. A multimodal agent (Claude Code on Opus/Sonnet, Codex on GPT models) inspects the template itself and builds a template_profile.json; text-only agents instead point the skill at a separate vision model via environment variables.
3. Approve the cover. The README's recommended flow previews one slide before the full deck, so style corrections cost one generation, not eight.
4. Let it run the full deck — the README documents 10-way concurrency by default, with a 10-page deck finishing in about 30 seconds — and collect the per-slide PNGs plus the packaged .pptx.
5. Iterate by targeted edits: say "change slide 3's subtitle" or "remove the footer", and the agent regenerates only the target slide through image-to-image editing, leaving the rest untouched.

For editable delivery, say so explicitly — the README's example asks for text and basic graphics directly editable, the hero visual independently movable, and the complete gpt-image-2 art direction preserved. The skill renders the finished -editable.pptx back to images for a page-by-page multimodal review, and it will not report an unreviewed editable file as a success.

Tips

• Match the task to the README's own fit table: title, subtitle, date, and footer edits are the most stable; metric cards work but every number needs checking; dense tables, financial reports, and legal long copy are the README's own "Weak" row ("Small text and numbers need strict human review").
• Real assets are preserved by default: product screenshots, logos, charts get embedded as independent source-image objects; reference-based redraws happen only when you explicitly allow them.
• If you only have a topic, examples/ ships starter slides_plan.md recipes for product launches, investor pitches, weekly reports, courseware, thesis defenses, and book talks.
• The template gallery evaluates whole presentations, not covers — new templates are tested across cover, section, content, and data slides against the incumbent before publishing.
• Related reading on this site: guizang-ppt-skill is an HTML-deck skill from op7418, whose separate NanoBanana-PPT-Skills repo this project started from, and image-to-editable-ppt-skill covers the editable-PPT workflow this project credits for ideas.

When Not to Use This

Skip it when the deliverable is text-native: the default mode produces full-slide images, so dense tables, financial figures, and legal long copy land in the README's own "Weak" row — small text and numbers need strict human review. Without a local PPTX renderer you lose editable mode, and template cloning can fall back to template-page images you export by hand. And if you need native, editable text by default rather than behind an explicit opt-in, a conventional template tool remains the better fit.

See the leaderboard for more skills.