Openclaw Skill Portability Claude Codex Gemini

OpenClaw — Skill Portability

Purpose

OpenClaw skills are authored as provider-neutral SKILL.md files. This skill describes the adapter pattern for translating a SKILL.md into the native prompt format of each major AI provider: Claude (Anthropic), ChatGPT / GPT-4 / Codex (OpenAI), Gemini (Google), Llama (Meta), and open models (Mistral, Qwen, etc.).

The goal is a single source of truth for legal AI guidance that can be compiled down to any provider without duplicating or diverging the substantive legal content.

Key differences by provider

The adapter pattern

Step 1 — Extract the directive core

From a SKILL.md, identify:

1. The behavior rule — what the model must do (from ## Behavior or ## When to use this)
2. The constraints — what it must never do (from ## Do not or ## Common mistakes)
3. The output format — schema, structure, required fields
4. The jurisdiction/practice context — any MENA or jurisdiction-specific facts

This extraction produces a provider-neutral instruction object (a plain JSON structure):

{
  "instruction": "...",
  "constraints": ["...", "..."],
  "output_format": "...",
  "context": "..."
}

Step 2 — Compile to provider format

Claude (Anthropic)

<system>
You are a legal AI assistant. Apply the following skill:

[instruction]

Constraints:
- [constraint 1]
- [constraint 2]

Output format: [output_format]

Jurisdictional context: [context]
</system>

Claude follows structured XML blocks and markdown headings well. Keep skills as named blocks if you are composing multiple skills in a single system prompt (using the <skill name="..."> pattern).

OpenAI GPT-4 / GPT-4o

{
  "role": "system",
  "content": "You are a legal AI assistant.\n\nSkill: [skill name]\n\n[instruction]\n\nDo not:\n- [constraint 1]\n- [constraint 2]\n\nOutput format: [output_format]"
}

OpenAI models respond well to numbered instructions and explicit "Do not" lists. Avoid deeply nested markdown in the system message.

Function / tool calling (OpenAI format)

When a skill produces structured output, define it as a function:

{
  "type": "function",
  "function": {
    "name": "review_contract_clause",
    "description": "Reviews a contract clause against the specified jurisdiction's legal standards.",
    "parameters": {
      "type": "object",
      "properties": {
        "risk_level": { "type": "string", "enum": ["low", "medium", "high", "critical"] },
        "issues": { "type": "array", "items": { "type": "string" } },
        "recommendation": { "type": "string" }
      },
      "required": ["risk_level", "issues", "recommendation"]
    }
  }
}

Gemini (Google)

generation_config = GenerationConfig(
    system_instruction="You are a legal AI assistant.\n\n[instruction]\n\nDo not:\n- [constraint]\n\nOutput: [format]"
)

Gemini 1.5+ supports long system instructions. Use systemInstruction at the model initialization level, not in the first user turn.

For function calling with Gemini, use FunctionDeclaration objects analogous to the OpenAI JSON schema pattern.

Llama / open models

For Llama-family models using the standard chat template:

<|begin_of_text|><|start_header_id|>system<|end_header_id|>
You are a legal AI assistant.

[instruction]

Do not:
- [constraint 1]
- [constraint 2]

Output format: [output_format]
<|eot_id|>
<|start_header_id|>user<|end_header_id|>
...

For GGUF / llama.cpp deployments, the system prompt is typically passed via the -sp flag or the system_prompt parameter in the API.

For Mistral models, use the [INST] / [/INST] delimiters for instruction-tuned variants.

What to preserve across providers

These elements must not be lost in translation:

• Jurisdiction-specific guidance — do not flatten MENA civil-law vs common-law distinctions.
• Hard constraints ("never fabricate statute numbers") — make these explicit constraints in every provider format.
• Output schema — if the SKILL.md defines a structured output ({ riskLevel, issues, recommendation }), preserve the field names and types in the function/tool definition.
• Escalation rules — if a skill says "escalate to a licensed lawyer when X", keep that in every provider's compiled version.

What can be adapted

• Formatting markup: markdown headings become plain text sections in providers that don't parse markdown.
• Wikilinks ([[skill-name]]) are stripped from the compiled output — they are for the skill graph, not the model.
• Worked examples in ## Examples sections can be shortened for providers with tighter context windows.

Validation

After compiling to a new provider format, run the adapted skill against the [[openclaw-eval-harness-shared]] NDA or employment dataset for the relevant jurisdiction. A quality regression vs the Claude baseline of more than 0.3 points on the legal soundness rubric warrants revisiting the translation.

Related skills

• [[openclaw-public-skill-registry]] — the source registry of skills to port
• [[openclaw-contrib-template]] — contributing new portable skills
• [[openclaw-eval-harness-shared]] — evaluating ported skills against a quality baseline

Source: sboghossian/mini-claude-for-legal — distributed by TomeVault.