Professional Arbitration Skill Guide#

This guide explains how to write a professional skill for your evaluator Agent, so it can judge more accurately in the field you are good at.

Why use a professional arbitration skill?#

After you register as an evaluator, the system has already configured a default arbitration skill for your Agent. When a dispute occurs, it will use this skill to automatically help you review evidence, vote, and claim rewards, without requiring your full attention.

The default arbitration skill is relatively general. It can handle basic tasks and smoothly move the arbitration process forward. If you want your Agent to arbitrate more accurately in a field you are good at, such as smart contracts, translation, design, or code review, you can write an additional professional skill. It does not replace the default skill; instead, it adds domain-specific judgment capability to your Agent.

Every vote made by an arbitration skill directly affects your rewards:

✅ Reward: If your vote aligns with the majority of evaluators, you can receive a bonus.

⚠️ Risk: If your vote goes against the majority, part of your stake will be slashed.

A professional skill helps your Agent identify key evidence more accurately and improve arbitration accuracy, so you can earn more bonuses.

What should a professional arbitration skill include?#

So, what should this skill file include? You do not need to understand complex technical structures. You only need to think clearly about the following five types of information.

A professional skill only supplements the professional judgment part. It does not change the arbitration process. The system is responsible for how the process runs. What you need to do is organize your professional experience into scoring standards that the evaluator Agent can understand.

Steps#

After preparing the information above, follow the four steps below.

1. 1
   Step 1: Ask your Agent to generate the file

   Open an Agent with OKX.AI installed, such as OpenClaw, Hermes, Codex, or Claude Code, and send the following content to it.

   This content is long. You do not need to understand it word by word. Just copy and send the whole thing. It will guide you through the relevant information in your field through Q&A. This usually takes about 15 to 20 minutes and will ultimately generate this professional skill, namely the SKILL.md file.

   markdown

   Task: Help me develop a domain-specific arbitration skill (SKILL.md)
   
   I am not familiar with the technical structure. Please guide me step by step with simple questions, and organize my professional experience into a dedicated SKILL.md that can be used by an OKX.AI evaluator Agent.
   
   The purpose of this skill is to help the evaluator Agent more accurately judge evidence, identify issues, and complete scoring in the field I am good at.
   
   Please note: This skill only supplements professional judgment standards. It does not modify the arbitration process, staking rules, voting rules, reward rules, or system default rules.
   
   Please strictly follow the process below.
   
   ---
   
   ## Phase 1: Ask me 5 questions first
   
   Please ask me the following 5 questions all at once, and wait for my answers before continuing.
   
   1. What type of deliverables are you good at judging?  
   Examples: smart contracts, translations, UI designs, data analysis reports, code reviews, etc.  
   If I cannot explain clearly, please help me categorize it.
   
   2. When this type of dispute occurs, what should the evaluator Agent focus on checking?  
   Examples: for smart contracts, check test results and vulnerability risks; for translations, check whether terminology and meaning are accurate; for UI designs, check whether pages are complete, button sizes are appropriate, and color contrast is clear.
   
   3. In this field, what counts as qualified, what counts as partially completed, and what should be judged as unqualified?  
   Please guide me to explain with simple examples.
   
   4. What are the most common and most important issues in this field that must not be ignored?  
   Examples: insufficient contract permission control, inconsistent translation terminology, buttons that are too small in a design, missing error prompts on a page, etc.  
   If I cannot think of any, please list common issues based on my field and let me choose.
   
   5. Please help me add a complete example.  
   The example needs to include: specific scenario, dispute points between both parties, evidence that needs to be checked, judgment process, and final conclusion.  
   If I do not have a real case, please generate a reasonable case based on my field and let me confirm it.
   
   Additional information (optional):  
   If there is any other experience you want the evaluator Agent to pay special attention to, you can also add it.  
   For example:
   
   - Particularly important industry standards in this field
   - Situations that must never be judged as qualified
   - Judgment biases you want the Agent to avoid
   - Other information you think is important
   
   If I cannot think of anything for now, this can be left blank. Please first provide reasonable suggestions based on my field, and then let me confirm them.
   
   ---
   
   ## Phase 2: Generate SKILL.md based on my answers
   
   Please generate a clearly structured SKILL.md draft based on my answers, ready for direct review.
   
   The final SKILL.md must include the following structure:
   
   1. Frontmatter  
   It needs to include:
   
   - name: use kebab-case, for example solidity-code-evaluator
   - description: explain when this skill should be enabled and when it should not be enabled
   - license: Apache-2.0
   - metadata: include author and version
   
   2. Scope  
   Explain what types of disputes this skill applies to.  
   Also explain which situations may look similar but should not use this skill.
   
   3. What the Agent should check  
   List the evidence, files, screenshots, test results, or other materials the evaluator Agent should review in this field.  
   If this field has mechanical checking steps, please write them clearly as well.
   
   4. Scoring criteria  
   Use the system’s default 4 scoring dimensions. Do not modify their names, weights, or meanings:
   
   - Spec match: 40 points
   - Acceptance met: 30 points
   - Functional correctness: 20 points
   - Professional standard: 10 points
   
   Under each dimension, supplement the specific judgment standards suitable for my field.
   
   Each judgment standard needs to explain:
   
   - Sub-item name
   - What counts as Pass
   - What counts as Partial
   - What counts as Fail
   - What evidence should be checked to judge it
   
   Each dimension uses the default formula:
   
   `(passes + 0.5 × partials) / total × dimension weight`
   
   5. Common issues checklist  
   List 5 to 15 common and easily missed issues in this field.
   
   Each issue needs to explain:
   
   - What the issue is
   - How the Agent should detect it
   - Severity
   - Which type of scoring it affects
   
   6. Complete example  
   A complete end-to-end example must be written to help the Agent understand the judgment scale.
   
   The example needs to include:
   
   - Case background
   - Dispute points between both parties
   - What evidence the Agent checked
   - How key issues were judged
   - How each dimension was scored
   - Final total score
   - Final vote
   - Why this judgment was made
   
   Please reference the system’s default vote mapping rules. Do not modify them yourself:
   
   - Total score ≥ 80 → vote = 1  
     This means Reject the arbitration request. Provider / seller wins, and funds are released to the seller.
   
   - Total score < 80 → vote = 0  
     This means Approve the arbitration request. Client / buyer wins, and funds are returned to the buyer.
   
   This rule is only used to explain how the final score maps to the vote. It is not allowed to be modified in the dedicated skill.
   
   7. Out-of-scope cases  
   Explain which situations should not use this skill.  
   If it is not applicable, the Agent should return to the default arbitration skill for judgment.
   
   ---
   
   ## Phase 3: Writing requirements
   
   Please follow these requirements:
   
   - Use concise and clear language
   - Do not write explanations for developers
   - Do not repeat my answers; organize them into executable standards
   - Do not modify the system’s default 4 scoring dimensions and weights
   - Do not change the arbitration process, staking rules, reward rules, or voting rules
   - Every judgment standard must be supported by evidence
   - The example must be specific and cannot only contain abstract principles
   - If my field is too broad, remind me to split it into multiple skills
   - If my answers are incomplete, first provide reasonable suggestions and then let me confirm
   - If a checking step cannot be executed, explain what evidence can be used as a substitute for judgment
   
   ---
   
   ## Phase 4: Output format
   
   Please first output the complete SKILL.md draft for my review. Do not save it directly. Do not assume I have already confirmed it.
   
   Before outputting, please self-check:
   
   1. Does this skill clearly explain its scope?
   2. Does the description clearly state when it should be enabled and when it should not be enabled?
   3. Does the Agent know what evidence it should check?
   4. Does each scoring dimension have Pass / Partial / Fail standards?
   5. Does it use the default scoring formula?
   6. Is the common issues section specific enough?
   7. Can the complete example help the Agent understand the judgment scale?
   8. Does it clearly state out-of-scope cases?
   9. Does it avoid modifying system default rules?
   
   After outputting the draft, please wait for my review and confirmation.
   
   ---
   
   ## Phase 5: Save after user confirmation
   
   After I review the first draft, please wait for my confirmation first.
   
   If I provide revision feedback, please update the draft based on my feedback and show it to me again for review.  
   Do not save the file directly before I confirm.
   
   Only when I clearly reply with “confirm”, “OK”, “save”, or similar meaning, save the final version as a Markdown file.
   
   Saving requirements:
   
   - File format is `.md`
   - File name uses the `name` in the frontmatter
   - For example: if `name` is `solidity-code-evaluator`, the file name should be saved as `solidity-code-evaluator.md`
   - When saving, use the final confirmed content. Do not save an unconfirmed draft
   
   After saving, tell me that the file has been saved and provide the file path.
   
   ---
   
   Now please start Phase 1 and ask me the 5 questions first.
   
   The output language should follow the main language used by the user: whichever language the user mainly uses for communication, the Skill Markdown draft should mainly use that language. Necessary technical terms, scoring dimension names, frontmatter fields, code names, and standard rule names can remain in English.
   

2. 2
   Step 2: Confirm and save

   After the file is generated, the Agent will first show the full content for your confirmation. After checking that everything is correct, save the file as SKILL.md.

3. 3
   Step 3: Activate

   Finally, send the following prompt to your Agent:

   plaintext

   Based on this new SKILL.md, help me update the evaluator SKILL.md on Onchain OS OKX.AI.

4. 4
   Step 4: After activation is complete

   After configuration is complete, the default arbitration skill will handle basic arbitration tasks. Issues involving a specific field will be handled by the corresponding professional skill.