Quick Test Drive: handsomestWei/patent-disclosure-skill for AI-Assisted Patent Work

handsomestWei/patent-disclosure-skill gained 38 GitHub stars today, and the reason is practical: it turns a general-purpose AI model into a focused assistant for Chinese patent workflows.

The skill covers invention, utility-model, and design-patent disclosure documents. It can help identify potentially patentable points, explain patent concepts in plain language, track relevant policy signals, and support responses to examination opinions. In other words, it provides a structured workflow rather than asking an LLM to “write a patent” from scratch.

For developers, the interesting part is that the skill can be paired with an OpenAI-compatible gateway and a model selected for long-form reasoning:

from openai import OpenAI

client = OpenAI(
    base_url="https://b-lost.com/v1",
    api_key="YOUR_B_LOST_API_KEY",
)

response = client.chat.completions.create(
    model="claude-fable-5",
    temperature=0.2,
    messages=[
        {
            "role": "system",
            "content": (
                "You are a patent-disclosure assistant. "
                "Separate technical effects, inventive points, "
                "prior-art risks, and missing implementation details."
            ),
        },
        {
            "role": "user",
            "content": "Analyze this product idea and draft a disclosure outline: ...",
        },
    ],
)

print(response.choices[0].message.content)

A sensible evaluation should measure time to first token (TTFT), total completion latency, cost per million tokens, and drafting accuracy against expert-reviewed samples. I would not publish benchmark numbers without controlling for prompt length, region, concurrency, and model routing.

B-Lost’s relay uses https://b-lost.com/v1, advertises 20% off list pricing, and supports native Anthropic /v1/messages prompt caching. For repeated patent templates, prior-art rules, and formatting instructions, caching could reduce cache-hit costs by 90% and improve consistency. The main caveat is validation: patent outputs still require qualified human review, especially for claim scope, novelty, and legal strategy.

Overall, this is a promising workflow skill—not a substitute for a patent professional, but a useful structure for turning raw engineering notes into reviewable disclosure material.