Paper2patent

Paper2Patent

Convert a paper into a Chinese invention patent application draft that a patent attorney (专利代理师) can review and file with minimal rework. The output is a draft, not legal advice; say so once in the delivery message.

What good looks like

A strong draft does four things the paper itself does not:

1. Claims that define a protectable scope. Claim 1 contains exactly the features needed to solve the technical problem (the distinguishing features over the closest prior art) and nothing paper-specific that a competitor could trivially design around. Specific values, model names and optional modules move to dependent claims and embodiments.
2. A specification that enables and supports. A skilled person can reproduce the invention from the 具体实施方式 alone, and every claim term is explained there. For AI inventions this means the model's modules/layers/connections, training steps and parameters, and how inputs and outputs map to the application scene (专利审查指南, 2026-01-01 revision).
3. Technical character. The solution is framed as technical means solving a technical problem with a technical effect, so it is not rejected as 智力活动的规则和方法 (专利法第25条).
4. Complete fidelity. Every technical feature is traceable to the paper. Patent-style generalisation is allowed where the paper supports it; invention is not. Missing information becomes an explicit 【待补充：…】 placeholder, never a guess — added matter cannot be supported later and misrepresents the inventors' work.

Workflow

1. Intake

• Pick the mode: direct (default: draft everything, mark gaps), human-in-loop (confirm disclosure status, invention points and claim plan with the user before drafting), text-only (no files), claims-only, or review (check an existing draft against the paper).
• Read the whole paper: every section, equation, figure, caption, table and appendix. For a PDF, render the pages that contain the method figures and look at them; for text-only input, work from the figure descriptions and captions. The drawings must follow the paper's actual structure. See references/input-requirements.md for PDF handling and the minimum input.
• Establish the disclosure status (arXiv? accepted/published? public code? talk?) and dates. This decides whether the patent can still be novel. Read references/patentability-and-disclosure.md — every time, because the answer changes what you tell the user.

2. Source fact sheet (internal)

Before drafting, list the technical facts you will rely on, each with its location (§3.2, Eq. (4), Fig. 2, Table 3, Appendix B). Separate: (a) the paper's own contributions, (b) prior art the paper describes (related work, baselines), (c) experimental settings and results. This list is the backbone for fidelity, for support_map, and for the 背景技术 section. Do not show it to the user unless asked.

3. Invention mining and claim plan

Identify the closest prior art (usually the strongest baseline or the method the paper improves on), the distinguishing features, the technical problem they solve and the technical effect they produce. Then plan the claim tree before writing any claim: triage every candidate claim-1 feature (necessary / embodiment choice / interface detail), give each separable mechanism its own dependent claim (one fallback position per claim, never merged to save numbering), check that every branch of a generalised claim 1 is supported in the description, add the standard carrier claims, and record all of this in notes.claim_plan. Do not simply transcribe the paper's pipeline into claim 1. Read references/claims-drafting.md.

4. Draft the five parts

Follow references/patent-drafting-standard.md for the structure, content and language of each part, including the paper-to-patent mapping (what to keep, transform, or drop). Write in Simplified Chinese patent register: 本发明/本实施例, never 本文/我们/论文/作者. Keep technical clarifications (what a constraint or state actually means) in the text; move drafting remarks to the notes.

Then run the sufficiency audit (patent-drafting-standard.md §7): for every feature of claim 1 and the main dependent claims, check that the description states its data format, decision rule, failure handling and parameters. What the paper gives goes into the text; what it does not give goes into notes.sufficiency as a gap — never filled in by you. This audit is what produces a useful handover list; a draft that looks complete but hides its gaps is worse than one that names them.

5. Drawings

Design each figure as an explicit graph of nodes and edges taken from the paper's figures and method text. If the method checks, revises, retries or iterates, draw that loop (decision node + return edge) in the flowchart that illustrates claim 1, not just the success path. Read references/drawing-generation.md. A typical AI-method set: method flowchart (图1, also the 摘要附图), model/data-flow diagram redrawn from the paper's main figure, apparatus block diagram with reference numerals, electronic-device diagram.