GPT-Image2 PPT Generation Skill
Turn a Markdown outline into polished 16:9 PPT slides with OpenAI gpt-image-2, or clone any .pptx template's layout and swap in your own content.
Calls OpenAI's official gpt-image-2 API with the user's own key, generates images locally and packages them into a PPTX — no destructive operations, but incurs API cost.
- Depends on OpenAI's paid gpt-image-2 API — a multi-page deck incurs one image-generation call per page
What does this skill do, and when should you use it?
This is a PPT generation skill for agents such as Claude Code, Codex and OpenClaw. It feeds a Markdown outline (or slides_plan.) plus one visual style into OpenAI's official gpt-image-2 Images API, renders each page as a high-res 16:9 image, and packages the result into a .pptx. Instead of traditional template-filling, it lets the model paint every slide as a complete visual composition for consistent aesthetics. It ships 265 style templates (10 initial, 22 featured, 233 contributed by designer @Xiamu Lingzi / 夏目玲子) and can also clone any user-supplied .pptx template. Output is full-page-image PPT by default; a natively editable mode (off by default) rebuilds text, shapes and assets into real PowerPoint objects only when explicitly requested.
Reads a slides_plan.md outline and converts it to slides_plan.; compiles per-page prompts from a chosen style MD plus its .layouts. sidecar (or from a vision analysis of the user's .pptx template); calls the gpt-image-2 Images API per page with reference images and post-pasted real-asset slots; packages results into a 16:9 PPTX via python-pptx; and outputs per-page PNGs, prompts., metadata. and the PPTX. Template cloning uses render_template.py to render the template to PNGs, then a multimodal agent (or an optional external vision endpoint) produces a template_profile.. It also supports --edit for precise per-element edits, --rollback for version history, and --editable to produce a PPTX with native text, shapes, connectors and image layers.
- Founders and PMs who need investor pitch or product launch decks fast, going from a one-line request to a presentable deck
- Employees with a corporate template but no design skills who upload a .pptx and want new content in the same visual language
- Students preparing thesis-defense or final-year project presentations using the academic styles
- Claude Code developers who want courseware, weekly reports or book-sharing decks generated from a sentence in the terminal
- Users who explicitly need text editable in PowerPoint, activating editable mode for native text boxes and movable assets
- Users who must embed real product screenshots, logos or paper figures precisely into slides
What are this skill's strengths and limitations?
- Full-page visual generation keeps every slide aesthetically coherent, with 10-way concurrency finishing 10 pages in roughly 30 seconds
- 265-style library spanning tech, business, academic, education and fashion, with an online gallery for previewing and copying prompts
- Template cloning includes 1-page-1-layout reuse detection to avoid visibly repeated layouts
- Real assets are post-pasted as independent PPT picture objects at original fidelity — medical images and financial tables are not AI-redrawn
- Trackable and reversible: metadata. records slide_spec versions, enabling element-precise edits
- Clear secret-hygiene design: never recursively reads the caller project's .env
- Default output is full-page-image PPT — text and numbers are not directly editable in PowerPoint unless editable mode is enabled
- Editable mode and template cloning hard-depend on local PowerPoint COM / Keynote / LibreOffice; containers, HarmonyOS, Termux and unusual architectures may not work
- Direct API mode bills per image under gpt-image-2, outside ChatGPT subscriptions — full-deck cost must be budgeted
- The project itself flags dense tables, financial reports and long legal text as not to be promised without strict manual review
- Text-only agents need an extra VISION_* multimodal endpoint for template analysis, adding configuration overhead
- The --backend codex fallback takes 30-60s per page and relays parameters via natural language, with occasional failures
How do you install this skill?
Clone the repo and run the one-line installer: git clone [email protected]:JuneYaooo/gpt-image2-ppt-skills.git && cd gpt-image2-ppt-skills && bash install_as_skill.sh --target claude (installs to ~/.claude/skills/) or --target codex (installs to ~/.codex/skills/). Alternatively paste the repo URL to Claude Code or another agent and ask it to install. For direct API use, set OPENAI_BASE_URL, OPENAI_API_KEY and GPT_IMAGE_MODEL_NAME=gpt-image-2; template cloning and editable mode additionally require a local PowerPoint COM, Keynote or LibreOffice render backend.
How do you use this skill?
After installation, trigger it with natural language, e.g. 'Make me a 6-page deck about how AI changes content creation, generate the cover first for my review' or 'Here is company-template.pptx — mimic its layout for an 8-page product strategy deck'. The agent drafts slides_plan.md for confirmation, converts it to JSON, smoke-tests one page with --slides 1, then runs the full deck into outputs/<timestamp>/. To revise, say 'change the subtitle on page 3'; the script uses --edit to regenerate only that page.
How does this skill compare with similar options?
The README explicitly names op7418/NanoBanana-PPT-Skills as the upstream reference: the project originally swapped the image backend from Nano Banana Pro to OpenAI gpt-image-2, rewrote and added styles, then extended into template cloning, PPTX packaging and a structured style library. If you prefer the Nano Banana path, look at the upstream project; this repo's differentiators are gpt-image-2 rendering, template cloning and editable-PPTX reconstruction.