Hello everyone, I'm Miles. Lately, while writing content, organizing materials, making demos, and handling deliveries, Codex has become my most frequently used tool.
Today, I want to share the 10 skills that are currently most useful to me. From tracking news and creating covers to decoding books and turning images into editable PPTs, each has its own unique purpose. I've included specific examples and results below; if you like any of them, you can try them out directly.
1. follow-builders: Your Personal AI News Editor
If you follow dozens of AI creators and want to know what they've been up to, you usually have to scroll through feeds, find podcasts, and pick out the valuable parts from English content. follow-builders, created by Zara Zhang, organizes all of this into a concise briefing.
It tracks a group of AI researchers, founders, product managers, and engineers, summarizing their updates along with related blogs and podcasts. You can specify Chinese, English, or bilingual output, and tell it whether you care more about products, technology, or creation.
For a Chinese news briefing, a prompt like this is enough:
1Use follow-builders to organize a Chinese briefing for AI content creators.2Focus on new product use cases, real operation demos, and creator experiences.3Clearly state who said what and why it's worth reading for each item, keeping source links.4List your additional suggestions separately.
Below is the generated briefing:
Results of the Chinese briefing trial.
For example, in one entry, Zara suggests that product demos should retain real interface operations and the thinking behind them. Reading this, you already know how this update relates to making tutorials; if interested, you can follow the link for details.
The value of installing this is helping you maintain a readable follow list. For those tracking AI products long-term, looking at a fixed briefing like this is much easier than deciding who to search for every day. You can also use Codex's scheduled tasks for daily delivery; try one in a chat first to see if it suits your taste.
Installation: zarazhangrui/follow-builders
2. Punk Skill: Choose a Style to Create Distinctive Article Covers
Punk Skill is my core illustration tool for creation. Specifically, punk-cover reads an article or topic, organizes the visual based on the platform, aspect ratio, and chosen style, and then calls an image generation tool to complete the cover.
When creating content, the hardest part is often translating "what this article is about" into an image. If you just give an AI a title, it might just put out a bunch of computers, robots, and glowing lines. Punk's style library considers the subject, text, whitespace, and texture together, giving the image a clear direction once a style is selected.
Look at this "Building an Obsidian Knowledge Base from 0 to 1":
Layered paper, purple structures, and note cards turn the knowledge base theme into a visible space; the main title remains the first thing you see. The thickness and shadows of the paper also give the image a hand-built feel.
Switching to a pixel art style yields a completely different effect:
Both images are Punk Skill style examples.
Characters, maps, and props form a complete game scene. This creates a direct visual association for teaching tool upgrades, level-based learning, or building projects.
You can put your own article in:
1Use punk-cover to generate a cover for the following article.2Platform: X; Aspect Ratio: 5:2; Style: 3D Paper Cut Concept Poster.3Keep the main title within two lines, and use one main subject to express the core idea.4Please actually generate the image.56[Paste full article here]
It's worth keeping long-term because you can reuse favorite styles for a series, only changing the theme and subject, without having to explore a new visual direction for every post. Actual image generation requires Codex to call an image tool; check the Chinese titles in the cover before using.
Installation: adrianpunk/Punk-Skill
3. cangjie-skill: Making Course Methods Actually Work for You
cangjie-skill (Cangjie) extracts methods from course scripts, lecture transcripts, and tutorials to organize them into callable Skills. It handles "when to use it, what to check first, and how to judge," making it perfect for knowledge with clear steps that you want to follow repeatedly.
For example, a course on customer interviews might remind you: ask less about attitudes toward hypothetical features and more about specific recent experiences. This sentence is easy to skip in notes; once turned into a skill, it can be used to check interview questions.
Using a self-written interview teaching script as an example:
1Use cangjie-skill to turn this "Customer Interview Outline Check" course script into a Skill.2Let it judge question by question: what evidence can this question collect, and does it suggest an answer?3For inappropriate questions, provide a more specific way to ask.
Applying the compiled skill to a new question: "Do you like automatic reminders?" (Context: choosing between reimbursement reminders and progress queries). The result is as follows:
Results of calling the skill generated from self-written teaching materials.
This question only asks about the user's attitude toward reminders and cannot compare the two directions. The response changed the question to one that more easily elicits experiences, such as:
Please recall a recent reimbursement; how did it go from preparing materials to submission? Were there any parts that needed supplementing or re-processing?
After that submission, did you try to find out the processing progress? If so, why did you want to know at that time, and what method did you use?
These two questions look for actual troubles in the submission process and progress queries, providing material to judge what to build.
It's worth installing because the judgment standards you've learned can be directly applied to your next task. The more specific the input course, the more useful the compiled skill; materials with only a few slogans are hard to turn into reliable methods.
Installation: kangarooking/cangjie-skill
4. GordenImage2PPTX: Finally, Beautiful PPT Images Can Be Edited
AI generates a perfect PPT page, but you need to change two words in the title. At this point, the whole page is just one image; clicking anywhere selects the whole thing, making small changes a hassle.
GordenImage2PPTX specifically solves this: it processes the background, frames, icon decorations, and text separately, then combines them into a PowerPoint file.
First, look at this input image where the title, pyramid, icons, and right-side descriptions are all in one image:
After conversion, individual selection boxes appear around text and graphic elements:
Image source: Gorden Sun project example.
The most valuable thing to see here are these selection boxes. The title has its own text box, the descriptions on the right can be modified individually, and the extracted icons can be moved, scaled, or replaced separately. Adjustments that previously required regenerating a whole page can now be done back in PPT.
Give the task like this:
1Use GordenImage2PPTX to convert this slide image into an editable PPTX.2Make text into editable text boxes and extract icons separately.3For modules in the frame that need independent adjustment, please process them separately.4Try to maintain the layout and color scheme of the original image.
I recommend this to anyone who frequently uses AI to create courseware, proposals, and presentation pages. It fills the gap in editing capability often missing after image generation.
Note that "editable" depends on the object: text can be changed directly; independent images can be moved and replaced; charts might not revert to native data-driven charts, and complex frames might still be a single image. Installation also requires configuring the image processing capabilities required by the project.
Installation: GordenSun/GordenSuperPPTSkills · GordenImage2PPTX
5. book-to-skill: Decode a Book and Keep the Methods
book-to-skill can organize a book into callable reading skills: where chapters are, what important concepts mean, and which part to check for a specific problem.
If you just read a book once, a normal summary might suffice. Methodological or professional books that need repeated use are better suited for this. You don't have to re-explain the whole book every time; you can ask the AI to go back to specific chapters to see what the advice is based on.
Using "The Art of War" as an example, you can assign the task like this:
1Use book-to-skill to organize the full text of "The Art of War" into a reading skill.2Keep chapter entries according to the thirteen chapters, and organize core concepts, judgment methods, and common misconceptions.3When answering questions later, point out the relevant chapter first before giving advice.4Analogies in modern scenarios should be separated from the original meaning of the book.
It will leave behind chapter data, a glossary, and a cheat sheet. The real utility shows in the next question: "I only have 20 minutes for a book sharing session and want to cover the whole book. How should I arrange it?"
Results of calling the skill after generating reading skills.
The answer focuses the sharing on "what conditions to compare and what costs to consider before doing something," then locates "Laying Plans" and "Waging War." The session can use 2 minutes to pose the question, 6 minutes for conditions, 5 minutes for consumption, and the remaining time for discussion and explaining the boundaries of analogies.
This is much better than compressing each of the thirteen chapters into one sentence for a time-limited session. The content remains organized by chapter, but you can query it with your current problem. Note that the sharing arrangement is a pedagogical suggestion, not a claim that the original book was teaching public speaking.
Installation: virgiliojr94/book-to-skill
6. show-me: Turn Code Explanations into Understandable Diagrams
After Codex finishes modifying code, the explanation might contain dozens of filenames, component names, and call relationships, which are still hard to read. show-me organizes these relationships into flowcharts, component trees, call trees, or before-and-after comparisons, and can even generate HTML explanation pages when needed.
This is a Skill recommended by Matt Pocock, from HumanLayer, authored by Dexter Horthy. It's perfect for those who want to understand project structures or review code changes without chasing long explanations paragraph by paragraph.
The following case explains how reading positions are restored: after a page update, when to accept the saved position, when to wait, and when to cancel because the user scrolled manually.
HTML diagram case from HumanLayer.
The three boxes in the diagram represent Idle, Waiting, and Completed. Condition labels on arrows show when states switch; red paths mark cancellation, and green paths mark completion. Following the arrows, you can find the answer to "what happens if the user scrolls now?"
With the diagram, finding the corresponding logic in the code becomes much clearer. You can also directly say to a confusing answer:
1Use show-me to draw the explanation just given.2Label in Chinese: where it starts, what judgments it passes through, and the final result.3If code changes are involved, highlight the differences before and after.4For complex content, make it an openable HTML diagram.
It's worth installing because checking what AI did shouldn't be harder than having it do it. The point of the diagram is to help you see relationships; for specific code, it still needs to draw based on actual files.
Installation: humanlayer/skills · show-me
7. grill-me: Clarify Requirements Before Starting Work
grill-me is a requirement-probing skill by Matt Pocock. Give it a plan or idea, and it will find undecided parts that will affect future work, providing questions and suggestions, and continuing to probe after you answer.
It's especially useful when "I roughly know what I want to do, but can't state specific requirements." AI is good at filling in a seemingly complete plan, but decisions like login, pricing, and sharing methods should have actual reasons.
For example, wanting to build a book material download site:
1Use grill-me to help me figure out the requirements for a "Book Material Download Site."2First, ask questions that affect the production direction, explaining the differences different choices would make.3Give your suggestions for each question, and continue after I answer.
The first round of probing looks like this:
Results of the first round of probing.
First, it asks who it serves: workshop participants or all public readers? The former already has a shared background, so materials can directly complement the course; the latter might need introductions, categories, and browsing entries.
Then it asks what materials can be provided. Self-made handouts versus distributable materials determine what the first version of the site actually holds. Stopping at these key choices is more useful than discussing button colors prematurely.
You just need to answer and let it continue probing. Its value is bringing the decisions the plan would have defaulted for you out into the open.
Installation: mattpocock/skills · grill-me. Handle its dependency grilling during installation.
8. last30days: See How Others Use It Before Deciding What to Do
When preparing to try a new tool, asking AI for "useful tips" often yields generic advice. last30days first checks recent community discussions to see what people are trying, what problems they face, and which practices are worth further study.
It can conduct research around a tool, a topic, or a group of products, supporting sources like Reddit, X, and YouTube (depending on access). For those creating topics or looking for new ways to play, it's very practical: you can start with a broad question and turn the found clues into actionable prompts.
The project showcased a pet image case. The initial question was just "how are people using ChatGPT to make dog pictures," and the research found ideas like "turning dogs into humans." Following up on specific prompts revealed requirements: upload pet photos first, map fur color to hair color, and retain expressions and recognizable accessories.
Project showcase case: last30days handles research and prompts; human images generated by ChatGPT.
Comparing them, the fur color, the tongue-out expression, and the accessory on the chest all became clues for the character design. The most useful part of this example is turning a vague "wanting to make pet content" into a creative task with reference images and specific requirements.
You can switch to a topic you're interested in:
1Use last30days to research how people have been using ChatGPT to create pet images in the last 30 days.2Find playstyles with actual works and explain what photos are needed for each.3Pick one suitable for beginners and give me a Chinese prompt I can copy directly.4Keep reference sources and indicate platforms not searched.
The images above are existing project cases used to illustrate how it pushes research into creation; running it now will show playstyles that change over time.
Installation: mvanhorn/last30days-skill
9. Defuddle: Extract a Webpage into Clean Body Text
Defuddle does one simple thing: extracts the main content from a webpage and saves it as Markdown. Markdown is plain text with headings, lists, and links, which can be put directly into note-taking software or given to Codex for further reading.
It's perfect for those who frequently collect web materials. Selecting all and copying brings along navigation, buttons, and recommended reading; screenshots aren't easy to search or excerpt. Saving the body text provides a clean draft for later notes, translation, or comparison.
For example, NASA's moon science page has mission navigation, news entries, and many columns. Give it to Defuddle:
1Use defuddle to extract the main content of this webpage:2https://science.nasa.gov/moon/facts/34Save it as moon-facts.md, keeping subheadings, body text, and links.5Keep the original English text first; do not translate or summarize.
The extracted file looks like this when opened:
NASA webpage extraction results, opened in an editor.
Sections like "How the Moon Got Its Name," "Size and Distance," and "Orbit and Rotation" are all preserved, along with relevant links in the paragraphs. The large navigation blocks from the website are gone, leaving mainly the content to be read.
It saves the time spent reformatting after copying. Note that extraction quality depends on the webpage structure; this example still retained some image credits and relative links. For long-term archiving, just have Codex fix the links. View the extracted file.
Installation: kepano/obsidian-skills · defuddle
10. skill-creator: Turn Your Repeated Instructions into a Skill
skill-creator is a tool for creating and refining skills. You tell it the task, rules, and desired results, and it helps you organize the skill description; for parts that need fixed checks, you can even add scripts.
It's suitable for work you've already done repeatedly and know how to do well. For example, checking file links before delivering materials: what to check, what to skip, and what to report if a problem is found can all be clearly stated.
You can request it like this:
1Use skill-creator to create a Skill that checks local file links in Markdown.2Report the line number of the link, the target file, and whether the file exists.3Skip remote URLs, images, and code examples.4Only check, do not modify the original text. Prepare a few samples with different scenarios to verify it.
Applying the generated skill to a demo document yields these results:
Results of calling the generated link-checking skill.
In the table, not-created.md on line 4 is marked as missing, while other found files show as normal. You can go directly to that line to add the file or change the address, without having to click through everything yourself.
The utility of this example is small, but it's very practical when done repeatedly. By leaving the check objects, exclusion ranges, and output formats in a Skill, next time you just need to say "check this document" to follow the same set of requirements.
I suggest starting with small tasks with clear boundaries like this. A skill that can stably check files is already helpful; as you get used to it, you can add new rules.
Installation: openai/skills · skill-creator
How to Install
Pick one, send the corresponding installation address to Codex, along with the name of the Skill. A repository might contain multiple skills, so remember to name the one you want to install:
1Please install [Skill Name] from the address below and check its required dependencies.2After installation, perform a demo using the materials I provide.34Installation Address: [Paste the link above]
You don't have to install all ten at once. Pick one you can use immediately and put your own materials in; the effect will be much clearer than reading ten functional descriptions.
I'm Miles, an AI algorithm expert who transitioned from a big tech company to FDE. I've worked on algorithm R&D and optimization deployment, as well as corporate training and delivery. My X account reached 10,000 followers in 15 days and produced three long posts with millions of exposures in one week, two of which later exceeded 2 million.
Follow me @miles_mazy, let's grow together and make money together.
