How Patent‑Point Mining Works in the Patent‑Disclosure Sub‑Skill

The patent‑disclosure sub‑skill performs patent‑point mining through LLM‑driven prompts defined in its SKILL.md manifest, using a two‑step workflow of candidate generation and point fusion.

The patent‑disclosure sub‑skill in the handsomestWei/patent-disclosure-skill repository automates one of the most critical early stages of patent drafting: identifying and refining the core invention concept. Unlike traditional rule‑based systems, this implementation relies entirely on carefully engineered markdown prompts to guide a language model through structured reasoning.

Patent‑Point Mining Architecture

The mining process is orchestrated by skills/patent-disclosure/SKILL.md, which maps workflow steps to specific prompt files. For invention‑type patents, the manifest points to prompts/invention/patent_points_analyzer.md as the primary driver of point extraction.

The workflow divides into two distinct phases with clear handoffs between them.

Step 3: Candidate Patent Point Generation

The patent‑points‑analyzer prompt instructs the LLM to produce 3–5 candidate patent points, each structured with four mandatory components:

• Technical background – context establishing the problem domain
• Core innovation – the novel technical solution
• Differentiation from prior art – initial competitive positioning (validated in later steps)
• Feasibility notes – implementation considerations

The prompt enforces technical coherence while permitting "reasonable inference" where input data is incomplete. This allows the system to operate effectively even with partial invention descriptions.

Step 4: Point Fusion and Selection

Once candidates exist, the same prompt file drives fusion and selection:

1. Overlap analysis – the LLM identifies compatible elements across candidates
2. Combination innovation detection – synergies where multiple technical elements merge into novel solutions
3. Method‑plus‑system evaluation – determining whether the invention spans both process and apparatus claims

The selected title must satisfy a concrete‑ness constraint: it must be specific enough to ground downstream sections like block diagrams and flowcharts. The prompt explicitly warns against vague "system/module" constructions that would fail this test.

Output Strategy and Multi‑Draft Support

By default, the skill delivers one optimal patent disclosure. When users request multiple drafts, the workflow adapts:

• The LLM first produces outlines containing title plus key differentiators
• It then negotiates the writing order with the user before full generation

This approach prevents redundant computation while maintaining flexibility.

Support for Other Patent Types

The same architectural pattern extends to utility‑model and design patents through dedicated prompt files:

Patent Type	Prompt File
Invention	prompts/invention/patent_points_analyzer.md
Utility Model	prompts/utility_model/patent_points.md
Design	prompts/design/patent_points.md

Each shares the Step 3/Step 4 logical structure but incorporates type‑specific terminology and evaluation criteria.

Code Example: Skill Invocation

from skill_runner import run_skill

# Invoke patent-point mining as part of full disclosure generation

result = run_skill(
    skill_name="patent-disclosure",
    parameters={"type": "invention", "title": "Smart‑Thermal‑Regulation System"}
)

# Access the mined and selected patent points

print(result["patent_points"])

The returned structure contains the final selected point ready for downstream processing by disclosure_builder.md.

Prompt Structure Excerpt

## Step 3: Candidate Patent Points

- Generate 3–5 candidate points
- Each must include: technical background, innovation, prior‑art differentiation, feasibility

## Step 4: Fusion and Selection

- Analyze overlaps between candidates
- Highlight combination innovations where elements synergize
- Selected title must support downstream disclosure (see disclosure_builder.md §7.9)

Summary

• Prompt‑driven architecture: No Python logic handles mining directly; all reasoning occurs through LLM interpretation of markdown prompts
• Two‑phase workflow: Candidate generation (Step 3) precedes fusion/selection (Step 4) in patent_points_analyzer.md
• Quality gates: Concrete title requirements and feasibility checks ensure downstream usability
• Type coverage: Invention, utility‑model, and design patents share structural patterns through parallel prompt files
• Manifest coordination: SKILL.md serves as the single source of workflow truth, routing to appropriate prompts by patent type

Frequently Asked Questions

What makes patent‑point mining "mining" rather than simple generation?

Patent‑point mining implies extraction and refinement from underlying technical material, not ex nihilo creation. The prompt explicitly requires grounding in supplied facts and enforces structured output with differentiation analysis—treating the invention concept as discoverable through systematic prompting rather than arbitrary composition.

How does the skill handle incomplete invention descriptions?

The patent_points_analyzer.md prompt permits "reasonable inference" where data is missing. This allows candidate generation to proceed with partial inputs, though the quality of resulting points depends on inference accuracy. The fusion step in Step 4 subsequently filters or combines these candidates to compensate for individual weaknesses.

Why is the same prompt file used for both generation and fusion?

Consolidating Steps 3 and 4 in one file maintains contextual continuity. The LLM retains awareness of its initial candidate rationales when evaluating combinations, reducing inconsistency. This design also simplifies the SKILL.md manifest, which needs only one file mapping per major workflow phase.

Where does the selected patent point go next?

The chosen point feeds directly into disclosure_builder.md (referenced in Step 4 at §7.9), which constructs the full patent disclosure including technical drawings, claims scaffolding, and detailed description. The concrete title requirement exists precisely to enable this handoff.

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