How the Routing Mechanism Works in Patent-Disclosure-Skill: Intent-Based Dispatch Explained

The routing mechanism in handsomestWei/patent-disclosure-skill is a declarative intent-classification system defined in the root SKILL.md that maps Chinese trigger keywords to specific sub-skill entry points while enforcing strict package isolation and workflow boundaries.

The patent-disclosure-skill is a modular suite of six specialized sub-skills designed for patent workflow automation. According to the repository source code, the top-level SKILL.md acts as a centralized router that parses user input, identifies intent through keyword matching, and dispatches requests to the appropriate sub-skill handler without permitting cross-package tool invocation.

Intent Recognition and Keyword Mapping

The routing mechanism relies on hard-coded intent keywords defined in the routing table of SKILL.md. When a user submits a command, the router scans for specific Chinese trigger words that indicate which patent workflow to initiate.

The recognized intent keywords and their corresponding sub-skill entry points are:

• 交底书 → skills/patent-disclosure/SKILL.md
• 申请文件 → skills/patent-application/SKILL.md
• 著录检索 → skills/patent-search/SKILL.md
• 读专利 → skills/patent-reader/SKILL.md
• 审查答复 → skills/patent-oa/SKILL.md
• 政策简报 → skills/patent-exam-policy/SKILL.md

Each keyword acts as a direct pointer to the sub-skill's configuration file, allowing the Agent-Skills runtime to load the appropriate toolset and prompt definitions.

Package Isolation Enforcement

A critical guardrail in the routing mechanism is the prohibition of cross-package calls. The router enforces strict boundaries between sub-skills to prevent dependency contamination.

If a sub-skill requires a capability that exists in another package, it must use its own isolated copy of that tool. For example, each sub-skill directory contains its own tools/ subdirectory with redundant scripts rather than importing from sibling packages. This architectural decision ensures that updates to one sub-skill's tooling cannot inadvertently break workflows in another.

Special Case Routing Logic

The router implements explicit prioritization rules for edge cases that deviate from standard keyword matching.

When the input matches a patent number pattern (e.g., CN202310000000.X) or ends with a PDF extension, and the user's intent context suggests "读懂" (comprehension), the router bypasses the standard disclosure workflow and immediately dispatches to skills/patent-reader/SKILL.md.

Conversely, the router never auto-initiates the policy-brief (政策简报), OA reply (审查答复), or application (申请文件) flows from disclosure or reader contexts. These workflows require explicit user invocation to prevent accidental document generation.

Workflow Boundary Guardrails

The routing mechanism maintains strict separation between different search capabilities to prevent tool misuse.

During the disclosure search step, the router restricts execution to the lightweight "一词一页" (one-term-one-page) search implementation located within the patent-disclosure package. When executing a bibliographic search (著录检索), the router dispatches to the full-featured patent-search package. These two search pipelines are never mixed within a single user request.

Implementation Example

While the production implementation is driven by YAML declarations in SKILL.md, the routing logic operates conceptually as follows:

def route(user_input: str) -> str:
    # Intent detection via keyword mapping

    intents = {
        "交底书": "skills/patent-disclosure/SKILL.md",
        "申请文件": "skills/patent-application/SKILL.md",
        "著录检索": "skills/patent-search/SKILL.md",
        "读专利":   "skills/patent-reader/SKILL.md",
        "审查答复": "skills/patent-oa/SKILL.md",
        "政策简报": "skills/patent-exam-policy/SKILL.md",
    }

    for kw, entry in intents.items():
        if kw in user_input:
            # Explicit trigger validation for application flow

            if kw == "申请文件" and not _has_disclosure_dir(user_input):
                raise ValueError("Application flow requires a disclosed directory")
            return entry

    # Patent number or PDF detection bypass

    if _looks_like_patent_number(user_input) or _is_pdf(user_input):
        return "skills/patent-reader/SKILL.md"

    raise ValueError("Unable to determine routing")

The actual dispatch is handled by the Agent-Skills framework parsing the routing section of the root SKILL.md file.

Routing Configuration Reference

The following table summarizes the user commands and their routed destinations:

User Command	Routed Sub-Skill File	Purpose
交底书	skills/patent-disclosure/SKILL.md	Draft patent disclosure document
申请文件 /path/to/disclosure	skills/patent-application/SKILL.md	Generate application file set
著录检索	skills/patent-search/SKILL.md	Search CNIPA bibliographic records
读专利 my_patent.pdf	skills/patent-reader/SKILL.md	Parse patent and produce analysis
审查答复	skills/patent-oa/SKILL.md	Build office action reply
政策简报	skills/patent-exam-policy/SKILL.md	Generate examination policy brief

Summary

• The root SKILL.md functions as a declarative router that maps Chinese intent keywords to sub-skill entry points.
• Six primary intents are supported: disclosure, application, search, reader, OA reply, and policy brief.
• Strict package isolation prevents cross-directory tool sharing; each sub-skill maintains its own tools/ directory.
• Patent number and PDF inputs trigger automatic redirection to the reader sub-skill when comprehension is implied.
• Explicit invocation is required for OA, policy, and application workflows to prevent accidental execution.

Frequently Asked Questions

What automatically triggers the patent reader sub-skill?

The router automatically dispatches to skills/patent-reader/SKILL.md when the user input contains a string matching a patent number format (such as CN2023XXXXXXXX.X) or references a PDF file, provided the context indicates a goal of reading or comprehension ("读懂"). This bypasses the standard keyword-based routing.

Can sub-skills share utility scripts or tools?

No. The routing mechanism explicitly prohibits cross-package calls. Each sub-skill must use its own copy of tools located within its respective tools/ directory. This isolation ensures that modifications to one sub-skill's utilities do not create breaking changes in other workflows.

What is the difference between disclosure search and bibliographic search?

The disclosure search uses a lightweight "一词一页" implementation confined to the patent-disclosure package, designed for quick term lookups during disclosure drafting. The bibliographic search (著录检索) invokes the full-featured patent-search package with comprehensive database access. The router strictly separates these to prevent mixing lightweight and heavy search operations.

Why does the application flow require validation before routing?

When routing to skills/patent-application/SKILL.md, the router checks for the presence of a disclosed directory via _has_disclosure_dir(). This validation ensures that application file generation only occurs when prerequisite disclosure documents exist, preventing the creation of orphaned application drafts without underlying technical content.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}