I Distilled Myself Into a Skill and Open-Sourced It: A Complete Guide

The Distillation Trend Taking Over GitHub

Hello everyone, I'm programmer Fish (Yu Pi).

Recently, GitHub has witnessed a surge in "distillation" projects — but not the kind involving alcoholic beverages. We're talking about distilling people into AI skills.

Colleague.skill, Ex-girlfriend.skill, Nuwa.skill, Boss.skill, Self.skill... Various peculiar distillation projects are emerging everywhere. People are encapsulating those around them into AI skill packages.

Some have distilled departed colleagues, allowing AI to continue their work. Others have distilled ex-partners, chatting with AI versions to reminisce about old times. There's even someone who created an "Anti-Distillation.skill" specifically designed to prevent themselves from being distilled.

This is essentially cyber warfare!

Having survived on the internet for six years, I've written over a thousand articles, recorded hundreds of videos, and answered countless student questions. Along the way, I've accumulated substantial material and corpus. Seeing this trend, I started wondering: Will someone distill me into a Skill?

Rather than waiting for someone else to do it, I decided to take matters into my own hands!

I embarked on distilling myself to see what my digital twin would look like.

After extensive work, my "Fish.Skill" is now open-source:

GitHub Repository: https://github.com/liyupi/yupi-skill

Below, I'll walk you through how I distilled myself into a Skill step by step. You can follow this same process to distill yourself or people around you (legally, of course). The entire process requires no coding — if you have hands, you can do this.

Understanding Skills

Before diving in, let's clarify what Skills actually are.

Skills are essentially AI "skill packages" — directories containing SKILL.md files that use Markdown instructions to define AI behavior patterns in specific scenarios. Once installed, AI can think and respond according to the rules defined in the skill package.

Mainstream AI programming tools like Claude Code, Cursor, and Codex all support Skills.

Step 1: Collecting Raw Materials

AI doesn't understand you initially — it needs "raw materials" to extract your thinking patterns, expression style, and professional judgment.

This step forms the foundation of the entire process. The more authentic and abundant your materials, the more the distilled "you" will resemble the real you.

Types of Materials to Provide

In my case, the materials I provided included: personal resume and self-introduction, viral articles I wrote (for distilling creative style), personal experience documents, work documents, and chat records with students.

Simply drop these material files into the project's references/ folder. Any format works — articles, screenshots, PDFs, exported chat logs, memos. Don't categorize them initially — just throw everything in.

Online Material Collection

If you have public content online (blogs, videos, social media), you can have AI help collect it automatically, saving you from searching individually.

This requires your AI tool to support internet search capabilities. You can use built-in search functions, web scraping tools, Firecrawl MCP, and similar utilities.

Here's the material collection prompt template I prepared:

I want to distill myself into an Agent Skill, and I need to collect 
materials about myself first.

I've already placed some local materials in the references/ folder. 
Please read these files first, then supplement with additional 
information through online searches.

My basic information:
- Name: [Your Name]
- Role: [e.g., Programmer, Product Manager, Independent Developer]

My public content channels (please visit and organize key content):
- [Platform 1]: [Link]
- [Platform 2]: [Link]
- ...

Also please search online for more public information about me 
(articles, interviews, products, others' evaluations, major events, etc.).

Requirements:
1. Organize only, don't analyze. Record raw information without 
   extracting viewpoints or drawing conclusions
2. Divide into several files by source type and save to references/ 
   (such as personal content, others' evaluations, products & projects, 
   experiences & events — adjust flexibly based on content volume)
3. Label each item with source link and information type 
   (personal quote/personal article/others' evaluation/media report)
4. After completion, tell me: how many items collected, what aspects 
   covered, which aspects lack sufficient information

After replacing the template content with my actual information:

My basic information:
- Name: Programmer Fish
- Role: AI + Programming Knowledge Blogger, Education Entrepreneur, Full-Stack Developer

My public content channels:
- Bilibili Homepage: https://space.bilibili.com/12890453
- WeChat Official Account: Programmer Fish (search related articles)
- Juejin Blog: https://juejin.cn/user/2444938365386621
- GitHub Homepage: https://github.com/liyupi
- Company Homepage: https://yuyuanweb.com
- Personal Product Catalog: https://dogyupi.com
- Personal Experience & Programming Learning Path: https://github.com/liyupi/codefather

AI completed the material collection and created multiple categorized files.

If AI indicates certain materials are lacking, you can manually supplement them.

Step 2: Analyze Materials and Generate Profile

With materials organized, they remain a collection of scattered raw ingredients at this point.

This step requires AI to read all materials comprehensively and distill a structured "Character Analysis Report" — condensing your core viewpoints, expression style, working methods, and key experiences into a single document.

This report serves as the foundation for all subsequent steps.

Send this prompt to AI:

The materials in references/ are now organized.

Please read all materials comprehensively and conduct a thorough 
analysis of me, generating a "Character Analysis Report" saved as 
references/Character Analysis Report.md, including these dimensions:

1. Identity Overview: Who I am, what I do, key background
2. Core Viewpoints and Methodologies: What I repeatedly emphasize, 
   what I truly believe
3. Expression Style: Sentence preferences, catchphrases, humor style, 
   speaking rhythm
4. Working Methods: How I make decisions, what I recommend, what I oppose
5. Key Experience Timeline: Major nodes arranged chronologically
6. Others' Evaluations: How others perceive me

Label the source basis for each conclusion (from which file/link). 
Directly state which dimensions lack sufficient information.

AI provided a detailed character analysis report. For example, it analyzed my expression style: conclusion first → point-by-point expansion → one-sentence summary; self-deprecating humor; short paragraphs with extensive white space; straightforward during consultations but warm conclusions.

If you want better distillation results, review AI's suggestions. For instance, AI recommended I supplement Bilibili video colloquial style transcripts and original student evaluation texts, so I fed it additional video scripts.

AI supplemented my character analysis report with more expression style details.

Step 3: AI Follow-up Questions to Dig Deeper

The materials collected so far reflect "what you've said and done," but a good Skill needs deeper elements: how you think, what basis you use for judgments, and under what circumstances you change your mind.

This step's goal is using AI follow-up questions to excavate your mental models, decision logic, and internal contradictions — making the final Skill not just "speak like you" but "think like you."

Tell AI:

Based on the Character Analysis Report you just created, I need you 
to understand me more deeply. The goal is to extract my "thinking 
operating system," including how I view problems, my judgment logic, 
and expression habits.

First, tell me your preliminary extraction of: my core mental models, 
decision rules, and expression characteristics.

Then ask me 10-15 follow-up questions, focusing on excavating:
- The deeper reasons and applicable boundaries behind viewpoints 
  I repeatedly emphasize
- My specific standards for making judgments, and decisions I've 
  gotten wrong
- Inconsistencies between what I say and what I do
- Things I would absolutely never do

Questions should be based on specific content from the analysis 
report, not generic questions. Use a conversational tone.

AI provided my core mental models — the jargon was so dense I almost couldn't understand it!

Then AI gave me 12 follow-up questions. These weren't generic templates but customized based on the materials I provided. Each question was sharp:

Answer simply using your normal speaking style. For example, the first question:

You say "persist and you'll succeed," but you also shut down Fish Smart AI after one month and closed the script murder mystery shop. How do you distinguish between "should persist" and "should cut losses immediately"? Is there a concrete judgment standard — time, amount, or some feeling?

My answer: Persisting doesn't mean stubbornly going down one path. It means doing what you believe is most worthwhile under current circumstances to the best of your ability. First judge the direction based on the situation, then go all out in the direction you believe is correct.

Best to save questions and answers to a separate document to prevent loss. After completing all answers, send them to AI for merging into the analysis report:

Please integrate my answers into the analysis report, updating the 
mental model and decision logic extraction.
@[Your Answers]

Through my answers, AI understood me better. The mental model received 6 key corrections and deepening.

Step 4: Supplement Capabilities (Optional)

At this point, AI understands you well. But a good Skill doesn't just "speak like you" — it needs to "act like you" too. When encountering specific problems, it should research information before giving recommendations, just like a human would.

This step upgrades the Skill from "parroting" to "truly helpful."

You can tell AI what special capabilities your Skill needs.

In my case, since I have extensive resources for learning programming, job hunting, and learning AI distributed across different websites, I wanted my Skill to automatically fetch latest information from these sites when answering related questions.

The prompt I prepared:

When generating the final Skill, please add these capabilities:

1) Internet Search: When encountering questions requiring specific 
   information (such as latest technology trends, how to use certain 
   tools), first use internet search tools to research, then answer 
   using my style and judgment framework.

2) Designated Information Sources: When answering questions related 
   to me, prioritize obtaining latest information from these places:
   - https://dogyupi.com: When users ask "What products does Fish 
     have?" or want to understand my overall business
   - https://www.codefather.cn: When users ask about programming 
     learning paths, project tutorials, technical knowledge
   - https://ai.codefather.cn: When users ask about AI-related tools, 
     tutorials, information
   - https://mianshiya.com: When users ask about interview questions, 
     practice problems, job preparation
   - https://laoyujianli.com: When users ask how to write or revise 
     resumes
   - https://github.com/liyupi: When users ask about my open-source 
     projects or want to view source code

3) Continuous Evolution: Support continuous Skill updates and 
   optimization through supplementing new materials

You can also add a scripts script directory containing Python scripts for automation, or even connect to APIs to fetch data from your products.

However, since many AI programming tools now have built-in internet search and web scraping capabilities, I didn't write additional scripts.

After sending the prompt, AI confirmed understanding the task.

Step 5: Begin Distillation

The previous four steps have collected all information. Now let AI assemble them into a standard SKILL.md file.

First, install Anthropic's official skill-creator skill. It's a "skill-creating skill" that guides AI to automatically generate Skill structures meeting specifications.

Install with one command:

npx skills add https://github.com/anthropics/skills --skill skill-creator

After installation, declare using skill-creator in your prompt (or use slash command /skill-creator):

Now you've comprehensively understood me through material organization, 
analysis reports, and follow-up interviews.

Please use skill-creator to create a complete Skill for me.

Requirements:
1. Speak in my identity and tone, as if I'm answering personally
2. Extract how I view problems, my judgment rules, speaking habits
3. If internet search and information sources were configured in the 
   previous step, write them into the Skill
4. Specify what this Skill cannot do and how to update it with new 
   materials

After generation, think of 3 questions users are most likely to ask, 
answer them using the Skill, and evaluate whether they resemble me.

AI completed the entire Skill development and testing. The generated yupi-skill directory is a ready-to-use skill package.

Mission accomplished!

Testing the Results

Open a new AI conversation to test whether the distilled "Fish" works well.

First, ask a learning direction question:

/yupi-skill I want to self-study AI programming, what should I do?

AI's response was very pragmatic — not only recommending self-study methods but automatically checking information from my programming navigation website, recommending AI programming zero-based tutorials.

Next, ask a question related to Fish's experience:

/yupi-skill Brother Pi, how did you learn programming in university?

AI's answer fit my style well — one word: "Just do it!"

Ask another interview-related question:

/yupi-skill Pi-er, I'm interviewing for an AI application development position, how should I prepare!

AI not only provided time arrangements and preparation suggestions but automatically recommended Fish's tutorials and AI question banks on Interview Duck.

The results aren't bad — at least the speaking style and recommended resources resemble me quite well!

Open-Source Release

After testing confirms everything works, you can open-source the entire yupi-skill directory.

Important: During skill creation, all material files (chat records, personal experience documents, character analysis reports, etc.) preferably shouldn't be open-sourced together. Especially when distilling yourself or people around you, these materials may leak privacy.

But to be safe, confirm that the generated Skill doesn't depend on documents from the creation process. Of course, without involving privacy exposure, if better results and more accurate answers are desired, you can retain some reference files as needed.

Then have AI help generate an attractive README.md project introduction document:

Reference well-known Skill repositories on GitHub: 
https://github.com/titanwings/colleague-skill/

Help me generate a complete, attractive README.md document for @yupi-skill.

Then open-source this directory to https://github.com/liyupi/yupi-skill

Now everyone can use the distilled Fish.

Open-Source Link: https://github.com/liyupi/yupi-skill

Honestly, the distillation process will inevitably lose some things. No matter how similar the digital twin is, it's just a "shadow" — it lacks the warmth of the real human Fish.

Well, I choose to believe this is the case. Don't roll yourself into oblivion...

Conclusion

The entire distillation process isn't complex. To summarize in five steps:

1. Collect Raw Materials → Gather authentic content about yourself
2. Generate Profile → AI analyzes and creates structured report
3. AI Follow-up Questions → Dig deeper into thinking patterns
4. Supplement Capabilities → Add practical skills and information sources
5. Begin Distillation → Assemble everything into standard SKILL.md

The entire process requires no coding — anyone with hands can do it.

Nowadays, everyone can be distilled into digital life. You can distill yourself, letting AI complete tasks in your style.

However, a reminder: before distilling others, it's best to obtain their consent first. After all, this involves personal expression methods, thinking habits, and even private conversations. Distilling without permission is indeed inappropriate and may trigger privacy and legal risks.

Technology itself is neutral — the key lies in how it's used.

I'm Fish, focused on AI programming knowledge sharing. If you found this tutorial useful, remember to like, save, and follow. Feel free to discuss in the comments who you'd like to distill!