Built a Skill for Openclaw to realize "building a self-evolving Agent from scratch"

The starting point of the matter: a problem that gave me a headache#

Recently, while playing with OpenClaw, every time I wanted to create a new Agent, I had to manually create a bunch of files: SOUL.md, IDENTITY.md, AGENTS.md, MEMORY.md…

None of them can be omitted, each has its own specific format requirements, and after writing them, I still have to run a bunch of CLI commands to register, bind channels, and install skills.

The whole process is tedious and error-prone.

One day, it suddenly occurred to me: Can I let AI help me with this?

Not just asking it to help write a single file, but rather — me saying “help me create an Agent” and it getting everything done.

Thus, there is the story to be told in this article.

What is a Skill?#

In OpenClaw, a Skill is an “operation manual” for AI.

You put it into the system, and when a user triggers the corresponding keywords, the AI will read this manual and execute step by step according to the instructions inside.

Essentially, a Skill is an automated process written in natural language.

What we need to do is write a Skill called create-agent, so that after reading it, the AI knows how to create a complete Agent from scratch.

Design Idea: Clarify the Process First#

Before starting, I went through the entire process with Claude.

Creating a complete Agent requires at least the following steps:

1. Collect information — Know what kind of Agent to build
2. Check for conflicts — Confirm that the ID is not occupied
3. Confirm channel parameters — Authentication parameters vary completely across different platforms (Feishu/Telegram/Discord)
4. Generate Workspace files — 6-8 Markdown configuration files
5. Execute CLI registration — openclaw agents add, set-identity, bind
6. Install evolution Skill — Enable the Agent to have self-learning capabilities
7. Output user to-do items — Tell humans what else needs to be done

Each step is indispensable, and the order cannot be messed up.

The Most Interesting Part: Self-Evolution#

This is the part I am most satisfied with in the entire design.

The created Agent is not a static robot, but one that continues to evolve with use.

How is this specifically achieved? Through two Skills that work together:

self-improving-agent acts as an observer. After each conversation ends, it records what happened in memory/YYYY-MM-DD.md in a structured format — what was done, whether it was successful or failed, what mistakes were made, and what was learned.

capability-evolver acts as the execution engine. It regularly scans the memory/ directory and .learnings/ logs, extracts signals from them, and generates evolution plans.

A one-sentence summary of their division of labor:

self-improving-agent is the diary, and capability-evolver is the engineer who reads the diary and then starts to transform itself.

Key detail: memory/YYYY-MM-DD.md is the connection point between the two. If the Agent does not write to this file in the specified format, capability-evolver cannot read the signals, and evolution will not be triggered. Therefore, we have fixed the format requirements in the AGENTS.md template.

In addition, all changes involving core configurations (SOUL.md, AGENTS.md) must be approved by the user before execution and will not be modified automatically randomly.#

A Pitfall Encountered: Channel Parameters#

When initially designing the “binding channel” step, I simply had users fill in something like telegram:ops and considered it done.

Claude directly pointed out the problem: The authentication parameters for different channels are completely different.

• Telegram only requires a Bot Token
• Feishu requires App ID + App Secret + Verification Token + Encrypt Key
• WhatsApp doesn’t need a Token but requires scanning a QR code
• Slack requires two Tokens (Bot Token + App Token)

Hardcoding a single field makes it impossible to handle such differences.

The solution was to create a new references/channel-params.md file, which includes parameter templates for 11 mainstream channels. When the Skill is executed, it identifies the type of channel filled in by the user, dynamically reads the corresponding parameter template, and then generates a questionnaire for the user to fill out.

For adding new channels later, you just need to add an entry to this file, and the main process remains completely unchanged.

There’s another detail: if the channel filled in by the user is the same as the channel where the current conversation is taking place, additional confirmation is required — whether to bind the current robot application or switch to a new one? If this isn’t clarified, it’s easy to mess up the configuration.

Final File Structure#

1
create-agent/
2
├── SKILL.md                        ← Main process (5 phases + 2 intermediate checks)
3
├── scripts/
4
│   └── generate-workspace.sh       ← Automatically generates all workspace files
5
└── references/
6
    ├── workspace-templates.md      ← Complete templates for 6-8 configuration files
7
    └── channel-params.md           ← Authentication parameter mapping table for 11 channels

The execution flow of the entire Skill is:

1
Phase 1   One-time questionnaire (required + optional)
2
Phase 1.5 ID conflict check (execute openclaw agents list)
3
Phase 1.8 Channel parameter confirmation (identify channel type → dynamic questionnaire)
4
Phase 2   Generate Workspace files (call generate-workspace.sh)
5
Phase 3   Execute CLI registration (add → set-identity → bind)
6
Phase 4   Install evolutionary Skills (self-improving-agent + capability-evolver)
7
Phase 5   Output user operation list

Usage Effect#

Say the trigger phrase “Help me create an Agent”, and the AI will:

1. Display a questionnaire where you only need to fill in the Agent ID, name, and personality theme; the rest have default values.
2. Automatically check for ID conflicts.
3. Display the corresponding parameter questionnaire based on the channel type you filled in.
4. Call the script to generate all workspace files.
5. Run all CLI commands in sequence to complete the registration.
6. Install the Evolution Skill.
7. Finally, provide you with a checklist telling you what else needs to be done (such as git init, scanning the QR code, etc.).

There are only two places where manual intervention is required in the entire process: filling out the questionnaire and confirming channel parameters; all other steps are fully automated.

#

A Further Extension#

After finishing create-agent, I took the initiative to develop create-agent-network.

It addresses a more complex problem: When you need not just a single Agent, but a collaborative network consisting of a main Agent + multiple sub-Agents + several Skills, how to automate the setup?

The process becomes:

1. Analyze the workflow in the user’s description (number of nodes, what each node does)
2. Determine whether each node is suitable for an Agent or a Skill
3. First, search ClaWHub to see if there are existing Skills available
4. Create each node in sequence (call create-agent or skill-creator)
5. Automatically register the routing relationships between the main and sub-Agents
6. Output the topology map of the entire network and the user’s to-do list

All parameters of the sub-Agents (ID, name, personality, routing conditions) are automatically generated based on workflow analysis, without requiring the user to fill them in one by one.

Final Thoughts#

What struck me most during this process was not “how much time AI saved me”, but:

When you start using AI to build AI infrastructure, the entire way of working changes.

Previously, I needed to understand the format of each configuration file, the parameters of each CLI command, the authentication method for each channel… Now, this knowledge is沉淀 into Skills, transformed into reusable process assets.

The next person who wants to create an Agent can directly use this Skill without needing to know anything else.

This is probably what is meant by “automation of knowledge”.

Already listed on Clawhub#

The link is:https://clawhub.ai/dao24dao/create-agent-arch Welcome to use and provide feedback.

Permanent Reminder, Don’t Miss It#

WARNING

Yangqing Blog: http://dao24dao.online

Free Planet: [https://t.zsxq.com/QxWxg](https://t.zsxq.com/Q

Welcome Gift#

1. Add Yang Qing’s personal WeChat (dao24dao), note “Official Account”, and receive: [Feiba Super Individual, Liang Kao Pu’s Road from 0 to 10 Million, findyi Super Individual 0 to 1 Million Monetization] e-book of the million-dollar side hustle secrets.

2. The latest electronic version of “Solo Enterprise Compound Interest Commercialization” can be obtained directly (for free) by replying 2501 in the background of the “Yang Qing AI Solo Enterprise” official account.