Markdown
Actually, before I set up this blog, I had never used Markdown (MD) markup syntax. The Mix Space blog framework was my introduction to MD.
When I later tried to build a wiki, a website, and MD-generated manuals and other practical examples, MD consistently served as a key content source, and it's very easy to do version control and diffing with tools like git.
Coincidentally, my deeper use of Markdown lined up perfectly with the AI field's shift from early "prompt" conversations to today's "skill + agent" as a genuine assistant. If you want to truly use AI well these days, Markdown and Markdown files are indispensable.
My boss and I have always had a huge disagreement: I think "content matters, and form matters even more," while he thinks form isn't worth worrying about at all.
No matter what the content is, it ultimately has to be read and understood by some "entity." Good form greatly improves readability; bad form can easily distort the message.
Markdown happens to make both writing content and adjusting formatting fairly simple, and it is easy for humans, AI, or any other script/tool to understand. In the modern world, perhaps all content writing should begin in Markdown and then be converted into the target form using tools or AI.
Agent
The progression has moved from directly calling a large model[^direct-llm] >>> compound AI systems[^compound-ai]. Now, especially after Lobster blew up, genuine agents[^agent] are emerging at full speed. Claude Code and GitHub Copilot have both become true agents.
In using agents, Markdown becomes even more important:
- As skill prompts, telling the agent the key tech stacks to focus on while working.
- As boundary conditions, setting the file boundaries the agent can operate on autonomously and which operations always require approval.
- As style preferences, whether to prioritize performance, or encapsulation and code reusability.
- As trusted sources, adding fully reliable reference material for the agent to consult.
- As persistent information, letting the AI reduce context, record the project's current state and project-specific priorities, and keep things consistent across project stages and environments.
Either way, even creating a skill.md with just one line written as if you were directly talking to a large model will still have an effect.
**你是一个xxx技术专家,擅长xxxx领域应用**
Learning About Agents
This didn't start because I wanted to learn about agents. The models that SillyTavern used were so bad that I couldn't play properly at all, so I had no choice but to deploy my own instance—and the experience was really good.
To get a better gaming experience, I wrote my own character/world cards, but my ideas were too complicated, so I had no choice but to deploy an agent to help with creation.
I copied this project and made some tweaks with Claude.
This project is relatively simple, but its mechanisms suit my taste very well. For details, see the project's development docs or the author's statement on the plagiarizing project (that part is written even more clearly than the original project's docs). I plan to build all of my personal agents based on this project.
This project also uses SOUL.md/MEMORY.md for overall skill definition and persistence. It also supports a local-first knowledge base, and knowledge documents should be stored in Markdown to make them easier for agents to understand.
Text Games Rise Again
AI still can't be cheaply integrated into modern games, but it is already shining in the old-school genre of text games. As someone obsessed with text games, I'm already having a blast.
All of these user-configurable parts of AI are still in Markdown. I think that when models become cheap enough to integrate into modern games, it will still be the same.
AI makes Markdown great again?
- direct-llm: Throw a prompt at the model, and it directly outputs a block of text.Returndirect-llm
- compound-ai: Usually in the form of a chat window. It can't take a long-term goal and keep iterating until it's done, but it does have some ability to preprocess and call tools, so you don't have to type out all sorts of prompts yourself.Returncompound-ai
- agent: It can autonomously call tools and decide what to do on its own, without you spelling out every step. It can run for a long time to accomplish each goal it is given and operate on files on your hard drive.Returnagent