# gpt-image2-ppt-skills: Image-First PowerPoint Generation and Template Cloning

> gpt-image2-ppt-skills generates 16:9 presentations as full-slide images with OpenAI gpt-image-2, then packages them into PPTX files. It suits users who prioritize visual consistency and template imitation, while its optional reconstruction mode addresses selected editing needs at the cost of extra rendering and processing work.

**JuneYaooo/gpt-image2-ppt-skills** — Clone any .pptx into your own deck — OpenAI gpt-image-2 mimics the layout, you supply the content. 10 bundled styles. | 把任何 .pptx 模板"抄"成你的 PPT：gpt-image-2 仿版式、你换内容，另含 10 套精选风格。Claude Code / OpenClaw skill. 

- Repository: https://github.com/JuneYaooo/gpt-image2-ppt-skills
- Website: https://slidecraft-template-gallery.vercel.app/
- Stars: 1,385 · Forks: 78
- Language: Python
- License: Apache-2.0
- Published: 2026-10-16 · Updated: 2026-10-16 · Language: en
- Canonical page: https://hysenlabs.com/projects/juneyaooo-gpt-image2-ppt-skills

## Why gpt-image2-ppt-skills Uses Full-Slide Images

gpt-image2-ppt-skills addresses a specific gap between image generation and presentation delivery. An agent can ask gpt-image-2 to design each page as a complete visual composition, after which the skill returns high-resolution PNG pages and a packaged 16:9 `.pptx`. The README calls this image-first behavior a departure from filling text into a conventional template. That choice gives the image model control over typography, hierarchy, illustration, spacing, and the relationship between text and graphics.

The intended users are people working through Claude Code, Codex, OpenClaw, Cursor, Trae, Hermes Agent, or another agent that supports Skills. The project is especially relevant to product introductions, investor presentations, training decks, courseware, thesis defenses, weekly reports, and reports that need to follow an existing company template. The repository also contains `slides_plan.md` examples for several of these scenarios.

This is a presentation-generation skill, not a general-purpose PowerPoint editing library. Its central output is a designed slide image. In the default mode, changing a title or number normally means asking the agent to regenerate the affected visual page through image-to-image generation. The README describes page PNGs, version tracking, rollback support, and targeted regeneration, but it does not document a conventional object-level editing API for ordinary use.

## Template Cloning and the gpt-image2-ppt-skills Data Flow

Template cloning starts with a `.pptx` file or an image representing a reference page. According to the README, the agent examines the template's layout, color palette, and visual language, then asks gpt-image-2 to create new content in that style. This is imitation rather than literal reuse of PowerPoint placeholders. The demonstration describes preserving the reference page's layout, colors, and illustration vocabulary while replacing it with material about a different subject.

The repository layout indicates a Python implementation under `scripts/`, supported by `requirements.txt`. The declared dependencies connect several parts of the workflow: `openai` handles the image model, `python-pptx` handles PPTX assembly, `pillow` and `pymupdf` support image and document processing, while `requests`, `python-dotenv`, and `jsonschema` cover API access, environment loading, and structured configuration.

The skill also supports a preflight based on rendering. Users are advised to generate one cover page for approval before producing the full deck. The README states that the default generation process uses ten concurrent image requests, but that statement is a project claim, not an independently measured result here. The correct expectation is a designed cover, then a full set of PNG pages and a PPTX assembled for display.

## Installing the Skill for Claude Code or Codex

The easiest documented installation path asks an agent to install the repository from GitHub and then restart. For manual installation, clone the repository and run the supplied installer for the agent you use. The command below installs the skill for Claude Code.

```bash
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
```

The README documents `--target codex` as the alternative. The next block shows that command exactly:

```bash
bash install_as_skill.sh --target codex
```

The documented destinations are `~/.claude/skills/gpt-image2-ppt-skills/` for Claude Code and `~/.codex/skills/gpt-image2-ppt-skills/` for Codex. After installation, send a natural-language request to the agent. For a first real use, request a six-page presentation and ask for a cover preview. For template cloning, upload a `.pptx` and specify the topic, page count, and the requirement to retain the visual language.

## Configuring OpenAI Access and Optional Vision Analysis

The skill can use the OpenAI Images API or an OpenAI-compatible endpoint. The documented configuration uses `OPENAI_BASE_URL`, `OPENAI_API_KEY`, `GPT_IMAGE_MODEL_NAME`, and `GPT_IMAGE_QUALITY`. The repository's `.env.example` gives the default endpoint and model, but the README recommends injecting credentials through the agent framework or system environment rather than committing secrets to a project `.env`.

```bash
OPENAI_BASE_URL=https://api.openai.com    # 或任意 OpenAI 兼容中转
OPENAI_API_KEY=sk-...                     # 必需
GPT_IMAGE_MODEL_NAME=gpt-image-2
GPT_IMAGE_QUALITY=high                    # low / medium / hi
```

The `.env.example` also defines optional `VISION_BASE_URL`, `VISION_API_KEY`, and `VISION_MODEL_NAME`. These are needed when the calling agent cannot inspect rendered `.pptx` template screenshots itself. A multimodal agent can analyze the screenshots and provide `template_profile.json`; a text-only agent needs a vision-capable provider. The repository does not establish a recommended endpoint default, so the operator must supply an endpoint and key they trust.

Python 3.8 or later is indicated by the repository badge, and `requirements.txt` declares `requests>=2.31`, `python-dotenv>=1.0`, `python-pptx>=1.0`, `jsonschema>=4.0`, `pymupdf>=1.24`, `pillow>=10.0`, and `openai>=1.83,<2`. The README excerpt does not show a separate virtual-environment creation command, so one should not be presented as part of the documented installation.

## Editable Mode Separates Text, Shapes, and Image Assets

Editable mode is explicitly opt-in. The model still creates a complete visual page first, then the skill reconstructs known text, basic shapes, connectors, and complex assets into PowerPoint objects. The README gives a specific example containing five native text objects, six native shapes, one clean plate, and one independently movable image asset. That example demonstrates a partial reconstruction rather than a guarantee that every visible element becomes editable.

The workflow depends on local PPTX rendering. Windows PowerPoint, macOS Keynote, or LibreOffice must be available. After generation, the skill renders the editable file to `editable_renders/page-XX.png` and uses multimodal AI to inspect the result. The README says that an editable file is not marked successful when no rendering backend is available. It also describes a controlled repair process involving deterministic fixes, pixel extraction, masking, and limited AI redraws before a whole-page regeneration is considered.

The main trade-off is fidelity versus native editability. Image-first mode gives gpt-image-2 a unified visual canvas. Editable mode introduces reconstruction and rendering checks, so it can require more processing and does not guarantee that complex illustrations, small text, or unusual layouts will be cleanly separated. The README specifically recommends human verification for data pages and discourages direct promises for dense tables, financial reports, or long legal text.

## Where gpt-image2-ppt-skills Is the Wrong Tool

The main limitation is the mismatch between visual reproduction and native document structure. A PPTX assembled from complete slide images can look consistent, but its text and graphics are not ordinary PowerPoint objects in the default mode. If a team needs to reuse slide masters, maintain hundreds of linked placeholders, audit every text box, or make frequent structural edits without image regeneration, a native PowerPoint authoring workflow is a better fit.

The project also depends on access to an image model and, for some paths, a separate vision model. An agent without image generation support cannot use the main generation path as documented. Local rendering is another operational boundary for editable mode. PowerPoint, Keynote, or LibreOffice must be executable, and the output must be rendered for inspection before the skill treats it as complete.

A real alternative is a conventional PowerPoint generation library such as `python-pptx`. That approach creates native text boxes, shapes, and layouts directly, which gives the author control over document structure and editing. It does not give gpt-image-2 the same freedom to compose an entire page as one visual artwork. The choice is therefore between image-first presentation design and object-first presentation construction. This project favors the former, with selected reconstruction added as an optional second stage.

## Maintenance, Releases, and Upgrade Cost

The repository is not archived, and its last push was on 2026-08-22. No GitHub releases are published, so users cannot rely on a named release cadence from the available repository information. The project is distributed as a Skill repository with scripts, examples, styles, documentation, and installer targets rather than as a versioned standalone application in the documented data.

Upgrading may therefore mean replacing the installed Skill files or rerunning the installation procedure after reviewing repository changes. The README does not provide a migration guide, compatibility matrix, or rollback command for the skill installation itself. It does describe page-level version tracking and rollback within generated presentation work, but that is separate from managing the installed repository.

The license is Apache-2.0. That identifies the project's declared licensing terms, but it is not legal advice. Users should inspect `LICENSE` and any bundled example or template files before redistributing generated assets or incorporating the skill into a commercial workflow. External image-model usage also depends on the terms and costs of the configured API provider, which the repository does not define.

## Conclusion

Adopt gpt-image2-ppt-skills when an AI coding or agent environment already supports OpenAI image generation and visual consistency matters more than native PowerPoint editing. Do not adopt it as a drop-in replacement for a conventional editable presentation builder when dense tables, exact corporate masters, or pixel-perfect object control are primary requirements. Verify first that the target agent can install the skill, the required image API configuration is available, and Windows PowerPoint, macOS Keynote, or LibreOffice can render editable output on the local machine.

## FAQ

### Does gpt-image2-ppt-skills create editable PowerPoint text by default?

No. The default output is an image-first deck with complete slide visuals. Editable mode must be requested explicitly, and it requires local PPTX rendering support.

### What rendering backend is required for editable mode?

The README identifies Windows PowerPoint, macOS Keynote, or LibreOffice as the required local rendering options. Without an available backend, the skill does not mark the editable file as successful.

### Can gpt-image2-ppt-skills use an existing PowerPoint template?

Yes. The template-cloning workflow accepts a `.pptx` or an image and uses its layout, colors, and visual language as references for new content. It imitates the visual structure rather than documenting a literal placeholder replacement system.

## Sources

- [Issues](https://github.com/JuneYaooo/gpt-image2-ppt-skills/issues)
- [JuneYaooo/gpt-image2-ppt-skills on GitHub](https://github.com/JuneYaooo/gpt-image2-ppt-skills)
- [License: Apache-2.0](https://github.com/JuneYaooo/gpt-image2-ppt-skills/blob/main/LICENSE)
- [Project website](https://slidecraft-template-gallery.vercel.app/)
- [README](https://github.com/JuneYaooo/gpt-image2-ppt-skills/blob/main/README.md)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/juneyaooo-gpt-image2-ppt-skills