AI Native Slides: A Native Slide Workflow Framework Designed for AI Agents

AI Native Slides is a meta-skill framework aimed at helping AI Agents select the correct workflow and representation when creating slides. It avoids premature commitment to specific formats (such as directly generating PPTX) and supports the complete iterative process from story ideation to final delivery. Its core philosophy is: most AI slide failures are essentially workflow failures. Through structured routing selection and quality check mechanisms, this framework helps Agents avoid common pitfalls and deliver useful slide outcomes.

When current AI creates slides, most Agents directly jump to generating PPTX. This premature format commitment easily leads to failure. Slide creation needs to go through stages like story ideation, evidence organization, visual direction determination, editable source file creation, and rendering quality check. Without proper workflow guidance, Agents tend to make the following mistakes:

• Generate slides before arguments are clear, resulting in empty content
• Call image slides 'editable', confusing previews with source files
• Prematurely choose PPTX, getting stuck in XML and layout issues
• Turn interview pitches into dense project checklists
• Deliver directly without rendering quality checks

AI Native Slides is not a simple generation tool but a meta-skill—guiding Agents to choose appropriate representations at each stage. The framework defines clear workflow stages and routing options:

Route	Applicable Scenarios	Notes
Feishu Docs + lark-cli	Story, notes, evidence, materials are still changing	Not suitable for final presentations
Slidev	Technical demos, web previews, Git-friendly editing	Requires layout adjustments
Beamer	Academic/PDF-first presentations	Not suitable for visually rich pitches
GPT-image → image-first	Quick visual direction and mood boards	Text embedded in pixels cannot be edited
Editable PPTX rebuild	Final PowerPoint handover	Slower but maintainable
PPTX / OpenXML	Final compatibility requirements	Heavy for early iterations

This mechanism allows Agents to choose tools on demand instead of using PPTX一刀切 (one-size-fits-all).

The framework follows three key design principles:

1. Separate Preview and Source Files: Slides generated by GPT images can quickly provide visual direction, but text embedded in pixels cannot be edited. It is necessary to clearly distinguish between previews and editable source files.
2. Editable Priority: For slides that need iteration, the editable rebuild process is recommended. Although slower than one-time image generation, the resulting PPTX can be edited by humans, making it suitable for collaborative scenarios.
3. Rendering Quality Check: Includes a dedicated rendering QA checklist to verify visual quality and content integrity before delivery, avoiding problematic slides.

The framework provides real demonstration examples to show the output effects of different routes:

• Feishu Docs document creation and access records
• Slidev technical demo export
• Beamer academic slides
• GPT-image quick visual prototype
• PPTX and OpenXML final format output

These examples help users understand the advantages and disadvantages of each route and make informed choices.

The entry point of AI Native Slides is the SKILL.md file, which Agents can read to get detailed instructions. Users only need to tell the Agent:

Install AI Native Slides from https://github.com/OpenClaudex/ai-native-slides and use it when you need to create/migrate slides, select routes, work based on image prototypes, perform editable rebuilds, or conduct rendering QA.

This design can be easily integrated into various Agent frameworks like Codex.

Currently, the project is in the 0.2-alpha stage, marked as 'Codex Skill Ready'. Future development directions:

• Expand support for more slide tools and formats
• Enhance the automation of routing selection
• Add more practical application cases
• Improve the rendering QA checklist

AI Native Slides represents a more mature AI-assisted content creation method. It recognizes that slide creation is a complex process that needs to consider workflow, representation, and iteration needs. Through structured routing selection and quality checks, it helps Agents avoid common pitfalls and deliver truly useful slide outcomes. For users and developers who use AI to create presentations, this framework provides valuable references and is worth learning and applying.