YouMind
تسجيل الدخول

I Spent 190M Tokens Building an Open-Source PPT Skill

@zhouluobo
الصينية09 أكتوبر 2026
120K
451
78
47
1.0K

ليرة تركية؛ د

The author released a free, open-source AI skill capable of generating and beautifying PowerPoint presentations. It preserves original text while applying diverse visual styles and supports direct editing in PowerPoint.

Hi everyone, I'm Luobo.

I usually make talking-head videos and write WeChat articles, often needing to whip up a PPT to go with them. Friends in my groups frequently ask if there's a tool that can turn an article directly into a presentable PPT.

Actually, I wrote an article before introducing a skill that could generate PPTs and smoothly beautify existing ones. The results were impressive, but the downside was that it was a bit expensive, making it hard for many friends to use.

https://x.com/zhouluobo/status/2104859645823103245

There were loud calls in the comments for a free version. So during this holiday, I burned through over 100 million tokens and finally built a free, open-source skill. Whether you're making a new PPT or beautifying an old one, the results are excellent. If you pair it with WorkBuddy, you can even save on LLM token costs. It's amazing.

zhouluobo - inline image

Words are less effective than showing the output. Let's look at the images.

First, see the effects: A picture is worth a thousand words

A standard PPT: Text unchanged, just a new outfit

These are the first few pages of a previous Muse script. The top row is the original; the bottom row is what it looks like after I gave it a mint-green reference image for beautification.

zhouluobo - inline image

Titles, bullet points, numbers—not a single character changed. Only the color scheme, layout, icons, and illustrations were replaced. That small green tip box in the original became a capsule bar with an icon at the bottom, instantly improving the look.

Same original, different reference image, completely different style

Using the same script, changing the reference image or simply naming a preset style swaps the entire theme. From top to bottom below: original, mint reference, dark tech, ink magazine.

zhouluobo - inline image

The text in all four versions is identical. You can use them for A/B testing or switch styles based on the occasion.

If the reference has windmills, the cover gets windmills

Just after National Day, I tested it with a report titled "Domestic Scenic Spot Market Observation." The reference image was an environmental science cover: mint green background, windmills on hills, solar panels next to small houses.

zhouluobo - inline image

The cover was changed to centered alignment, with a yellow brush stroke under the subtitle. The landscape at the bottom was drawn based on elements from the reference image. This didn't use any AI image generation model; the mountains, windmills, and houses were drawn programmatically line by line. So your program can do the same—even without configuration, you get this effect.

Science courseware works too, formulas intact

I also tried it with a high school math calculus review courseware. 11 pages, deep blue with orange accents, overall simple but monotonous layout.

zhouluobo - inline image

For this one, I didn't provide a reference image or specify a style. It kept the original deep blue and orange color scheme, added a faint grid paper background like math scratch paper. On the right side of the cover, it drew a function graph with tangent lines at point P and secant lines approaching it, matching the lesson on "limit of average rate of change." Formulas were placed in light bars, centered. The rules for determining increasing/decreasing intervals kept the original green and red colors.

The real challenge here was formulas. The original had 12 formulas created with a formula editor. When I opened it in LibreOffice, those spots were blank, indicating that such editors have high compatibility requirements.

zhouluobo - inline image

So in this skill, I required converting these formulas into plain text formulas, e.g., V′(x) = (a−2x)(a−6x). They are visible in any software and directly editable.

Turn an article directly into a presentable PPT

I fed this very article to it, asking for a 15-page talking-head PPT in ink magazine style. Here is the complete set it produced.

zhouluobo - inline image

It first generated a "ghost draft" with only titles. Just looking at the block below tells you what the whole presentation is about.

text
11. A PPT skill that can edit and generate
22. Table of Contents
33. What it can do
44. AI-made PPTs: either uneditable or uniform
55. Beautification changes only appearance, text untouched
66. One skill packed with so much
77. Why it's reliable
88. Why are AI-made PPTs recognizable at a glance?
99. Write conclusion titles first, then discuss aesthetics
1010. From just beautifying to generating and self-checking
1111. Compared to common practices, differences in four areas
1212. How to use
1313. Install in three steps, use with one sentence
1414. Turn this article into a 12-page talking-head PPT, ink magazine style.
1515. First run it on an old article

Each page title is a conclusion statement. For topic-style titles like "Data Analysis" or "Project Introduction," it marks them for you to change. The table of contents is auto-generated from three chapter pages, with linked numbering and styling. Every page has speaker notes prepared; open the PPT notes section to read along.

Same content, change style with one sentence

I asked it to create versions in "Blue Line Launch" and "Mint Fresh" styles for this article PPT. Each row below is a style, showing the same four pages.

zhouluobo - inline image

Table, timeline, and comparison page structures remained unchanged; only the visual language style changed. The cover in Ink Magazine is pure large text, while the other two include illustrations. Very flexible.

Data, processes, analysis frameworks fit too

Complex content isn't a problem. Timelines, funnels, radial centers, 2x2 matrices, native tables, native charts, plus SWOT, PEST, Porter's Five Forces, Business Model Canvas—all have ready-made layouts. No fear of overflow.

zhouluobo - inline image

Another cool thing: All above are PowerPoint native objects. Text is in text boxes, tables are real tables, bar/pie charts can be double-clicked to edit data, and charts update accordingly. Only backgrounds, illustrations, and icons are images.

14 styles, just pick a name

Without a reference image, you can choose from presets.

zhouluobo - inline image

Note that most differences between these styles aren't just colors.

Black Gold relies on flowing gold threads; Launch relies on concentric ripple lines; Performance Report uses torn-edge dark bands; Ink Magazine and Swiss Grid barely use decoration, relying entirely on typography. These decorations are drawn programmatically, so even without an AI image model, you get complete visuals like above.

Who did we learn from? Standing on giants' shoulders

This skill isn't entirely reinventing the wheel; I referenced many excellent projects.

zhouluobo - inline image

dashi-ppt-skill: Learn restraint for beauty

I dissected several PPT skills on GitHub, seeing how their SKILL.md files and prompts were organized. I learned the most from dashi-ppt-skill, which clearly specifies capacity limits for each layout (how many items, characters per item).

I think this approach is great for avoiding overflow issues, so I added a capacity feature to this skill.

Funnels max 5 layers, each description ≤26 chars; 2x2 matrix quadrants max 3 items, each ≤14 chars—all written as a table. During outline writing, it checks against this table and warns if exceeded. During layout, it automatically switches to a fitting layout.

Many AI PPTs shrink fonts because they don't know page capacity, forcing text in, ruining layout.

273 templates from design sites: The essence

I curated 273 templates from PPT design sites, categorized into 36 visual routes, each with sample images. I fed this material to AI to learn. Honestly, this is the core essence.

For example, almost every template has a TOC page after the cover, with chapter numbering matching the TOC style. At the end, the closing page echoes the cover's visual style.

Good styles have decorative motifs. Black Gold performance reviews rely on flowing gold threads; launch styles repeatedly use concentric ripples.

Writing these design rules into the skill immediately improved overall quality.

open-design: Checking is more important

open-design is an open-source project by nexu-io. I found useful methods for PPT from its design style and learned a set of layout checking rules.

For example, don't scatter accent colors everywhere; fix font sizes to specific levels. Each rule is quantifiable. The skill includes checking scripts: max 3 accent color uses per page, font sizes must come from a fixed ladder, etc.

It also has a five-dimension scoring method: Unity, Hierarchy, Detail, Communication, Rhythm. After each output, there's a self-evaluation step, ensuring product quality control.

zhouluobo - inline image

Testing in WorkBuddy reveals more issues

Finally, testing. Initially, I ran it in Claude Code. Due to different LLMs, some issues weren't caught early. So I ran cases in WorkBuddy and indeed found problems.

In short, building such skills requires repeated testing across different Agents and LLMs to ensure stability and robustness.

What makes it better than other PPT skills?

Here are my judgments, may not be perfect, take them with a grain of salt.

zhouluobo - inline image

The most practical point: Output is directly editable. If unsatisfied, tweak it yourself in PowerPoint; no need to beg AI to regenerate.

Beautification doesn't change text. It verifies character-by-character via scripts, not relying solely on LLM compliance, because LLMs aren't always reliable.

Works without AI image models. All images above were generated without connecting image models. With them (OpenAI, Gemini, Volcano Ark), illustrations become richer.

Install anywhere. Supports Claude, Codex, Cursor, WorkBuddy, and other skill-enabled Agents. Installation script detects installed Agents and handles dependencies.

Downside: Slow. A ~12-page deck takes 15-25 minutes. The WorkBuddy test took 45 minutes. Some optimizations exist, but time is a hard constraint—slow work yields fine results.

So I added a fast mode. If in a hurry, say "fast mode," it skips intermediate confirmations, producing output in ~10 minutes.

Take it and use it

If you use WorkBuddy, it's simple: Go to the skill market in the left menu, search "Luobo PPT," click install.

zhouluobo - inline image

You can also use the open-source version:

https://github.com/zhouwei713/luobo-ppt

Download locally, unzip, enter luobo-ppt folder, run installation script. First use checks environment automatically.

macOS and Linux

text
1bash install.sh

Windows: Double-click install.bat, or run in PowerShell

text
1.\install.ps1

To install for a specific Agent, add a parameter, e.g., bash install.sh --agent codex.

Or give the open-source URL to any Agent and let it install.

After installation, tell the Agent:

text
1Turn this article into a 12-page talking-head PPT, with speaker notes per page, ink magazine style.
text
1Help me beautify this PPT, following the style of this reference image, don't change a single word of text.
text
1Theme is "Annual Review of AI Tool Accounts," give me three style samples, I'll pick one to proceed.
text
1This PPT will be projected in a meeting room, use projection mode, larger fonts.

14 Style Reference Table, pick by name when choosing styles.

Final Thoughts

Honestly, this skill took huge effort. Beyond burning hundreds of millions of tokens, organizing hundreds of PPT templates from design sites took days.

But the result is good. Whether generating or beautifying PPTs, I'm satisfied. Now fully free and open-source for everyone, hoping for feedback.

Any usage issues or optimization suggestions, please share. Luobo will humbly accept.

Next step:

Find an old article you've written, run it through once.

Fix whatever looks off. Still off? Come find me.

ريمكس في YouMind

قم بتحويل مقال سريع الانتشار إلى سير عمل كامل المحتوى

قم بتجميع المصدر وفك تشفير النمط وإنشاء الأصول وصياغة القصة وتوزيعها من مساحة عمل واحدة تعمل بالذكاء الاصطناعي.

اكتشف YouMind
للمبدعين

حول Markdown إلى مقالة 𝕏 نظيفة

عندما تنشر كتاباتك الطويلة، فإن الصور والجداول وكتل التعليمات البرمجية تجعل تنسيق 𝕏 مؤلمًا. YouMind يحول مسودة Markdown كاملة إلى مقالة نظيفة وجاهزة للنشر 𝕏.

حاول Markdown إلى 𝕏

المزيد من الأنماط لفك التشفير

المقالات الفيروسية الأخيرة

استكشاف المزيد من المقالات الفيروسية