Author: TinTinAI
Imagine a scenario: you are in the supermarket's fruit section picking watermelons, casually opening your phone to ask Doubao, and within seconds it provides you with several criteria for distinguishing based on patterns, sounds, and the stem of the melon which you can use on the spot. In life, we have quietly gotten used to AI.
But in work, AI often disappoints. Your boss asks you to write an event copy, and the result is a pile of clichés like "empowering the industry, co-creating the future, opening a new chapter."
Where does the problem lie? It's not that AI can't do the work; it's that you haven't learned how to "delegate tasks." More accurately, the way to use AI has evolved into the "you state your needs, and it gets it done" Agent era.
On the evening of June 6, the first live session of TinTinAI invited Validere AI engineer and Ph.D. from a prestigious American university, Teacher Dazhuang. Starting from the basic concept of Agent, he walked through the installation of OpenClaw, configured Feishu, and then broke down real Skill cases, completely demonstrating how an efficient AI assistant is "crafted."
The following is a summary of the highlights from this live session. Students who missed the live broadcast can visit "TinTinAI" on Douyin or WeChat Video Account to watch the full replay.
Technical Evolution: From Language Model to Agent
The Essence of Language Models: Predicting the Next Word
Around 2020 or 2021, language models (LLM) were merely predictive tools in laboratories, functioning similarly to the autocorrect feature of a mobile input method. For example, when you type "today the weather is," the input method might predict and complete with "good," "bad," or "hot" based on probability.
Although this sparked discussions in the industry, the general public did not really use it because of its weak tool attributes.
The Birth of ChatGPT and the Extension of Abilities
It was not until ChatGPT emerged that scientists further trained and refined the model, allowing it to converse with you like a true partner. Because the model absorbed a massive amount of data during pre-training, people began to interact with it as a search engine to obtain all kinds of knowledge.
However, at this stage, AI still did not have "hands and feet," meaning it couldn't genuinely help humans execute and complete complex tasks.
Subsequently, OpenAI quickly added two groundbreaking features to ChatGPT:
Web Search: Allows AI to obtain the latest information from the internet in real-time, no longer relying entirely on static data from pre-training.
Function Call: The language model can answer questions in a dialogue box and also call already written external tools.
Once a language model possesses the abilities of internet connectivity and tool calling, it has taken the initial form of an intelligent agent (Agent).
What Are the Differences in Interaction Modes Between Agent and Traditional LLM?
To compare with software developed by programmers:
Traditional LLM: If a programmer encounters code that they can't write, they ask ChatGPT, which gives 1, 2, and 3 steps. The programmer executes step by step on their own, and when encountering a bug, they paste back the error for multiple rounds of iteration to achieve results. The human is the executor, and AI is the consultant, requiring a lot of manual information handling in between.
Agent: The programmer only needs to say "I want to develop this feature," and the Agent autonomously breaks down the steps, calls tools to execute, collects feedback, and adapts adjustments, solving the problem end-to-end. Core nodes still require human review, but overall efficiency is on a whole different level.
👉 In short: previous AI was like a "customer service" (I ask, you answer); now the Agent is akin to an "intern" (you help me do it, and tell me the result once done).
Skill: The Key to Generating Compound Interest Effects for Agent
After understanding Agent, a more core concept emerges: Skill.
This concept initially originated from the internal documentation of the Anthropic team's engineers.
They discovered that different engineers would generate very different prompts for the same task. Thus, they began to maintain a shared document internally, forming a standard workflow document that could be iterated over time, shared, referenced, and plugged in.
This is the embryo of Skill, which is essentially a "compound interest project."
When is Skill Not Needed?
Not all tasks are suitable for turning into Skills, and blindly developing them may create additional burdens. The following two types of scenarios do not require Skill:
Deterministic Programmatic Workflows: For instance, finance personnel reconcile accounts every month based on a cash flow table, which has explicit calculation formulas and standard processes; using Excel formulas or traditional workflows is more precise and better maintained.
One-off Temporary Tasks: A trivial task assigned by a leader on short notice, where interacting directly with ChatGPT or Doubao suffices. The time cost of developing a Skill far exceeds the cost of solving the task itself.
When is Skill Truly Suitable?
The most suitable scenario for Skill is when "input is non-standardized, and output needs to be standardized" in repetitive tasks.
A typical scenario is information gathering in the investment industry: each input source is different and can include various formats such as PDF financial reports, Excel spreadsheets, web articles, etc.
With Skill, AI will clean, process, and organize this data according to the standard process you set, ultimately outputting results in a uniform format.
Practical Demonstration: Building a Skill for Writing Weekly Reports from Scratch
Next, using the open-source Agent tool OpenClaw (colloquially known as "little lobster") as an example, I will demonstrate how to connect AI with Feishu to build an automated weekly report Agent.
Step 1: Tool Installation and Environment Configuration
Method A: Official Website Command Line Installation
Visit the OpenClaw official website and find the corresponding installation command for your operating system.
After installation, proceed to the configuration stage, with two core components to configure: the API for the language model and the channel for interaction with the Agent (such as Feishu).
Method B: One-click Lightweight Configuration on Feishu
Use the built-in one-click version of OpenClaw (openclaw.feishu.cn) on Feishu; just follow the prompts to confirm, suitable for users who do not want to go through the hassle of configuration, though at the cost of slightly lower flexibility and inability to switch language models easily.
💡 Beginner's Tip:
If you encounter issues during installation, you can use a third-party AI-assisted programming tool OpenCode (opencode.ai), describe the error in natural language, and it will help you debug, which can successfully solve the issue in the vast majority of cases.
Step 2: Initializing the Personality of the Agent and Native Skill
After connecting OpenClaw in Feishu, you can first do a character setup using natural language, telling it your desired style and address.
Next, have it list the existing native skills; it will display the basic capabilities that have been configured in Feishu one by one, such as "Feishu document read and write," "file management," "permission management," etc. These individual skills serve as the "building blocks" for assembling more complex workflows later.
Step 3: Building a Dedicated Weekly Report Skill
In the Feishu chat box, directly give OpenClaw the command: "Help me design a Skill that automatically generates Feishu weekly reports." It will proactively guide and ask you: what is the specific positioning of this Skill, and in what scenarios will it be triggered?
Once you describe it clearly in natural language, it will automatically generate and activate a new Skill in the background, typically containing the following components:
Skill Name: For example, feishu-weekly-report
Description/Trigger Conditions: Clearly specify under what specific command or context this particular workflow needs to be invoked.
Workflow (Pipeline): List detailed execution steps for each stage (e.g., read specified temporary fragment document -> extract unstructured information -> categorize and refine according to the established professional framework -> automatically create a document in Feishu and write a formatted weekly report).
Writing Principles: Define the tone of the output content (e.g., concise and professional, results-oriented, key data quantified, etc.).
The entire structure is automatically generated by AI all at once. After generation, you can also make detailed adjustments according to your personalized requirements for the report format at any time.
Step 4: One-click Trigger, End-to-End Delivery
To test the effectiveness, we can provide a very scattered, unstructured, and disorganized record document of daily operations for a self-media account (for example: “edited a video on Monday, replied to DMs on Tuesday”) to Feishu little lobster and say: “Using the Feishu weekly report Skill, help me summarize a weekly report based on the information provided in this document.”
It recognizes the trigger words, calls the Skill just created, requests document permissions, and automatically completes the entire process of reading, classifying, generating, creating, and writing, ultimately outputting a neatly formatted weekly report in Feishu along with item statistics for each category. The entire process requires no intervention from you, and you just need to review the final results.
Skill is not just a one-time creation that guarantees success; it needs to be continuously refined based on feedback—this is the significance of "compound interest projects": the more you use it, the more skillful it becomes, and the more it is polished, the more valuable it gets.
Can Start Immediately Without Knowing How to Code
Many people think they need to write code when they hear "configure Agent."
The entire demonstration process did not involve writing a single line of code, and all interactions with OpenClaw were completed through natural language. Even when faced with more complex tasks, you can use OpenCode to describe your requirements, and the code will be generated by AI itself.
The most valuable lesson from this session is not a specific tool—tools update too rapidly to keep up with.
What truly matters is a way of thinking: breaking down repetitive tasks into steps, crystallizing them into standardized Skills, and assigning them to Agents for execution. Understanding the basic principles allows you to combine them with your work scenarios, making you more proficient and efficient than others.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
