Hello everyone, I'm Miles. Nowadays, when I make PPTs, I no longer arrange them page by page myself. I let AI do it for me. This applies to my training sessions, school and corporate presentations, and even those state-owned enterprise documents with very detailed layout requirements.
Today, I've organized these methods, included the results, and will discuss how to use them in Codex, which parts save effort, and where you need to be extra careful.
I. Personal Speeches and Sharing
1. Slidev, Marp: Check these two first for courses and technical sharing
Both tools can turn Markdown content into presentations, making them particularly suitable for frequently updated courseware. You can let Codex modify a piece of text, replace code, or add a page to regenerate the slides, keeping the entire style consistent.
Slidev is better for interactive sharing. It supports code highlighting, formulas, flowcharts, step-by-step animations, and web components, allowing you to explain while demonstrating technology. Marp is suitable for conventional courseware dominated by text and images; content files and themes are separated, so subsequent revisions are direct without needing to redesign pages every time.
When using them, give the materials to Codex, specify Slidev or Marp, let it organize the content page by page, and then choose a theme, preview, and export. Since these are presentation tools themselves, you don't necessarily need a skill with the same name; Codex can call them via documentation, and community skills can supplement production standards.
Limitations mainly lie in complex design and delivery formats. Slidev has many features but also more project dependencies and interactive debugging. Marp's regular pages are easy, but doing a lot of free-form collage requires writing extra styles. Slidev's exported PPTX files are images, and Marp's are also image-based by default; experimental editable exports require extra dependencies and have limitations with complex styles and notes. If the other party needs to continue editing text and data, don't treat them as replacements for native PPTX.
Slidev Documentation · Slidev Export Instructions · Marp CLI
2. Guizang: guizang-ppt-skill
Guizang's advantage is its clear design rules. Titles, images, and body text have specific positions, and there's a distinction between chapter pages and content pages, making it easy to maintain consistency across a dozen pages. It offers two directions: electronic magazine and Swiss International Style, and considers speaker notes and presenter mode, suitable for live sharing.
First, choose a style, give the script to Codex, and let it match the cover, body, and chapter layouts. Try making one body page to confirm word count and image proportions before continuing with the whole set; if unsatisfied, change a specific layout category instead of rethinking every page.
The limitation is its specific design preference. Dense tables, long explanations, and strictly specified corporate templates might not fit well. The output is mainly HTML, which is convenient for browser-based presentations, but if you need to keep an editable PowerPoint or let others edit in Office, you'll need another solution.
Project Address: guizang-ppt-skill
3. Zhang Zala: frontend-slides
This one uses three color blocks for three viewpoints with minimal text, suitable for projecting on a screen to accompany an explanation. Zhang Zala's frontend-slides creates web presentations, with the advantage of being able to view different visual directions before choosing: fonts, color schemes, white space, and layout are adjusted together, offering more creative space when making a presentation from scratch.
In Codex, give it the sharing materials and preferred reference pages, specify frontend-slides, and do a style preview first. Once selected, create the body text and view it page by page in the browser. If a page is too crowded or an image is inappropriate, point out the specific page for modification. The Skill provides production rules, but the browser and export tools still need to be prepared according to project instructions.
It's suitable for those willing to spend a little time choosing samples and wanting to break away from fixed templates. Accordingly, style selection requires your feedback, and complex pages still need page-by-page inspection; external fonts and images also need offline playback considerations. Its strength is web design; native PPTX delivery is not the default result of this workflow.
Project Address: frontend-slides
II. Reports, Defenses, or Delivering Files to Clients
1. ppt-master: Retaining editable text and charts
These tasks prioritize whether the file can still be used after delivery. ppt-master designs the page first and then converts text, shapes, and other content into PowerPoint objects, allowing for both visual inspection and future editing. It also supports its own PPTX templates, notes, and native charts and tables, suitable for projects with clear delivery requirements.
Give the report, data, and reference styles to Codex, clarifying which content must be editable, especially tables and charts. Make one body page and one data page first, export them, and check if both can be modified normally before doing the whole set. It uses SVG as an intermediate page, so you can check the appearance first and then the converted PPTX.
Charts need to be more specific: draggable rectangles can form a bar chart, but they won't update data like native charts. You can find chart components and embedded workbooks in the project's native chart samples; during actual delivery, try changing a number to see if the graphic follows.
Its limitations are in conversion details: complex SVG effects aren't guaranteed to enter PowerPoint exactly as they are; fonts, animations, and charts all have support ranges. After selecting native chart mode, also check if the required type is supported. This path saves manual object creation, but layout checks in the target software are still necessary.
2. slide-skill: Presentations with clear themes and layouts
slide-skill also uses the SVG-to-PPTX approach, providing theme and template-related workflows. Its advantage is not having to start from a blank page; body, data, and conclusion pages can follow unified rules; after the source pages are kept, content changes can be regenerated. It's suitable for situations where the outline is clear and similar materials need to be produced repeatedly.
In Codex, choose a theme first, then provide page-by-page content and ask it to render a preview after generation. Public file samples have native text and shapes, but this only shows the file structure; specific pages still need to be checked for Chinese line breaks, font sizes, and charts.
The limitation is that ready-made themes might not suit complex brand designs, and charts converted into shapes might not retain data. I prioritize checking the page with the longest title; if phrases are split or body text is too small, I go back to that page to adjust. Just running the file doesn't mean the layout has passed.
3. deck-dna: Following existing company styles
If a company already has approved presentations, there's no need to choose a style every time. deck-dna extracts font sizes, colors, image/text positions, and text volume per page from two or three old drafts, then uses these rules for new pages. Its advantage is continuing truly used design habits, including how much text to put on a page and how to highlight data, which is more specific than just telling AI a brand color.
Give the old drafts and new content to Codex and let it process them this way. Don't just provide the cover for reference drafts; body and chart pages better explain the actual layout. Logos, footers, confidentiality markers, fixed color values, and other unchangeable parts should also be specified separately.
It depends on the quality of the reference drafts; if the provided styles conflict, the extracted rules will be vague. Learning the appearance doesn't guarantee retaining the original master slide; if the client requires a specific master slide or fixed placeholders, use that file directly. If you're just changing a few numbers every week, there's no need to relearn the style.
III. Science Popularization, Courseware, and Graphic Sharing
1. Baoyu: baoyu-slide-deck
The blackboard and watercolor examples show the difference in full-page image generation: background, brushstrokes, illustrations, and text together form the picture. Baoyu's baoyu-slide-deck is suitable for knowledge cards, science diagrams, and presentations sent for reading. It can turn abstract content into easier-to-view images; illustrations and backgrounds don't need to be spliced separately, making it great for courseware with strong visual styles.
In Codex, confirm the copy page by page first, then call the corresponding image generation tool. Generate one page to confirm the style, then try one body page to see if the text volume is appropriate, and then follow the same color, brushstroke, and proportion set. The required image generation tools or model access need to be configured separately.
The modification cost for this path is also clear: once text is drawn into the image, changing a sentence might require remaking the page. baoyu-slide-deck merges full-page images into the PPTX; titles in the file won't become text boxes because of this. Names, dates, numbers, and formulas must be checked against the original script.
Baoyu's other project, baoyu-design, follows HTML design, with the advantage of being able to adjust the layout in the browser first and then choose the corresponding export method, including editable PPTX export. It requires matching export tools and discrete slide page structures; you can't losslessly turn any webpage into a PPT. It's a different choice from full-page image generation.
baoyu-slide-deck Repository · baoyu-design
2. PPTAgent: Organizing presentations from papers and reports
If there's a lot of material, the difficulty is explaining the content clearly first; you can look at PPTAgent. It connects material analysis, planning, and generation, reducing the work of moving content back and forth between documents, outlines, and pages. It's suitable for sharing based on reports and papers where you need to organize the explanation order before designing pages.
This is a project that requires dependency configuration, not a single-file Skill you can use right away. You can let Codex assist in checking the environment and running it according to project documentation, using a piece of public material to test the workflow first. After generation, map each page's conclusion back to the original text, checking numbers, charts, and constraints.
Its barrier to entry is higher than a single layout Skill; parsers, models, and search dependencies all need configuration. Long materials might miss conditions, and speculations in the original text might be written too definitively, so sources must be kept and verified; for just a few pages of simple introduction, it might not be worth deploying the whole project.
IV. Already Have Image Drafts, Want to Change Text or Images
1. Gorden: Image2PPTX
After splitting a full-page image, text and some image elements can be operated separately. The most useful part of Gorden's Image2PPTX is preserving existing visuals, so you don't have to remake the whole page just to change a title. It's worth considering when you have generated images, old courseware screenshots, or only a PDF; it's especially suitable for small local modifications.
Give the original image to Codex, point out what you plan to change, and process one page first; PDFs can have the required pages exported first. It needs to recognize text, remove text to fill the background, separate elements, and then restore text boxes. After reconstruction, view the original image and the new page side-by-side to check for missing words, background repair marks, fonts, and line breaks.
Selecting a box doesn't mean everything has become a native graphic. Icons, frames, and artistic text might still be images—movable, but internal lines cannot be modified. If a page of data needs frequent updates in the future, remaking a native chart is usually more appropriate.
Project Address: GordenSuperPPTSkills
2. LRriver/AIPPT: Generating and splitting in a workbench
This project puts image generation and layered editing together, suitable for those who want to continue adjusting in an interface. After generation, you can continue to split text and move elements, reducing the back-and-forth of exporting and importing. The image above shows text and images being selected separately, suitable for cases needing both visual effects and local modification.
You can let Codex assist in running it according to project documentation, using one image to verify splitting and exporting before processing formal materials. The limitation is that generation and reconstruction capabilities depend on the connected services; complex illustrations won't automatically restore to vector objects; projects in the browser should also be exported and backed up promptly, as local storage doesn't mean automatic synchronization or permanent preservation.
Project Address: LRriver/AIPPT. This is not the same project as the AiPPT commercial service.
If you only have a PDF and need to restore text, you can also look at NBLM2PPTX: it focuses on separating text from the background, suitable for reassembling pages where text cannot be directly modified; complex charts might still remain as images, and font and background repairs also need checking.
V. Want to Put Generation, Modification, and Export Together
1. codex-slides: Change the outline first, then the page
codex-slides puts research, outline, style, and page editing into a workbench. This interface is very intuitive: change the title, points, and order first, then generate the page; if there's a problem with the content, you don't have to start over from the whole finished product. Each stage is visible, making it easy to return to a specific page for adjustment, suitable for those who frequently make presentations and don't want to manage the process solely through chat records.
You can let Codex assist in running it according to repository instructions, check dependencies, and try generation and export once. It leans toward image-based page production; the modification interface is convenient, but the output doesn't become fully native PPT because of it. For files requiring massive updates to table data later, I wouldn't make it my first choice; deployment and model access also require maintenance.
2. Presenton: Templates, editing, and model access
Presenton provides generation, editing, export, and model access, suitable for long-term use of a fixed production workflow. Templates keep the style of the entire draft consistent, model selection is more flexible, and it can also connect to other processes via API. Compared to one-off generation scripts, it's more suitable for use cases requiring repeated production and modification in an interface.
After letting Codex assist in setup, connect the model, choose a template, and use a simple material to verify editing and exporting. Limitations include deployment, model costs, and export compatibility, all of which you must handle yourself. Self-hosting doesn't mean offline; materials might still be sent out when connecting to cloud models; if you only occasionally make a PPT, you might not need to maintain such a set of applications.
Project Address: presenton/presenton
VI. If I Had to Choose, I'd Keep This Combination
Based on my usage of frequent teaching, training, and corporate sharing, I would keep two main workflows: Marp for daily courseware and ppt-master for formal delivery; image generation is added as an illustration tool. I only switch to Slidev when live interaction is needed, and image decomposition is reserved for when no source draft is available. This way, there are fewer choices, and commonly used styles can be repeatedly accumulated.
1. Courseware I present myself: Mainly Marp, Slidev as needed
For most courseware consisting of text, images, and step-by-step instructions, I use Marp first. Codex organizes the content and applies a fixed theme; for subsequent additions of knowledge points or case changes, I just keep modifying the source file. I only use Slidev if I need code demonstrations, formulas, step-by-step animations, or web interaction, as there's no need to increase debugging work for a few pages of static content.
The prerequisite for this choice is that I play it myself or deliver a PDF. If the organizer requires an editable PPTX, I go straight to the solution below to avoid having to convert formats right before the deadline.
2. Corporate sharing and formal delivery: Template + ppt-master + image generation
If there's a designated template, keep it, especially the fonts, footers, logos, and fixed layouts; only use deck-dna to extract rules if there are reference styles but no strict master slide requirements. Body text and charts are made into editable objects by ppt-master, and image generation is used only when the cover or chapter illustrations need a sense of visual impact.
I value the subsequent modifications of this combination more: if a client changes a title or a budget update adds a row, text and data can be handled separately without having to redraw the illustrations. It requires one more round of image-text assembly and export checking, but for files needing repeated modification, this cost is worth it.
Baoyu's full-page image generation is more suitable for making knowledge cards, science pages, or presentations where the content is set and rarely modified later. If there's already a good image and you just need to change the text, give it to Gorden; don't make "turning everything into images first, then splitting them all back" the default workflow.
3. Actual production only focuses on a few checkpoints
First, let Codex list the content page by page, delete empty talk and repetitive explanations, and then make the cover, a regular body page, and the most complex page. After these three pages are confirmed, batch generate; if a page can't fit, reduce content or split the page—don't just keep shrinking the font size.
Clean up the copy once before layout, and read it again after the finished product is out. Delete groundless terms like "significant improvement" or "comprehensive upgrade"; titles should state specific things; explanations for the stage go in the notes, while versions for independent reading should complete the context.
Before delivery, open it with the target software, change a title, replace an image, update a chart number, and play it through once. If fonts change, images are lost, or the page count is wrong, check the fonts, material paths, and export settings. If you need to present offline, try it here with the internet disconnected.
Finally, save the original materials, editable source drafts, assets, and exported files separately. Keep confirmed covers, body pages, and data pages as references for next time; continuing with the same set of methods is easier than picking a new Skill every time.
I am Miles, an AI algorithm expert who transitioned from big tech to FDE. I've done algorithm R&D and optimization deployment, as well as corporate training and landing delivery. My X account reached 10,000 followers in 15 days and wrote three long articles with millions of exposures in one week, two of which later exceeded 2 million.
Follow me @miles_mazy, grow together, make money together.
