How to use the Patent Drafting Pipeline agent skill

What is an agent skill, and what does this one do?

An agent skill is a folder containing a SKILL.md file with instructions that an AI agent such as Claude Code reads when a matching request comes in. This skill is one such folder. Its main file is SKILL.md, and it tells the agent how to run a fixed sequence of patent drafting steps.

The pipeline runs prior art search, novelty checking, invention structuring, claims drafting, specification writing, an examiner-style review, and jurisdiction formatting. It supports invention patents and, for China only, utility models. It also keeps a state file so a long draft can resume after interruption, and it waits for your confirmation at each checkpoint by default.

Install the patent-pipeline skill

Download the skill as a ZIP from the directory page and unzip it. You can install it in one of two ways: put it in your personal skills folder at ~/.claude/skills/, or in a single project's .claude/skills/ folder. Alternatively, in the Claude apps, upload the ZIP under Settings → Capabilities → Skills.

• Personal install: unzip into ~/.claude/skills/ so the skill is available in every project.
• Project install: unzip into the .claude/skills/ folder inside your project so it is available only there.
• Claude apps: upload the ZIP under Settings → Capabilities → Skills.

Check what the skill needs before you start

The skill needs Claude Code or a compatible agent that supports its allowed tools: Bash, file read and write, web search and web fetch. For the examiner-style review step it also uses the Codex MCP server (mcp__codex__codex), which needs a configured reviewer backend.

It can optionally read a .aris/reviewer.json reviewer configuration. It also chains sibling skills such as prior-art-search and claims-drafting, which should be installed in the same repository. Note that the skill runs scripts and uses the network.

Trigger the skill with a natural request

You do not need a special command. The skill is meant to be triggered by phrases such as 写专利, patent pipeline, 专利申请, draft patent, 写权利要求书, or a plain request to draft a complete patent application. Describe the invention and name the target jurisdiction in the same message.

The skill's argument hint is an invention description followed by a jurisdiction, separated by a dash. You can also override defaults in the request, for example asking for a utility model or for all three jurisdictions.

• Draft a patent for my sensor calibration method, target US.
• patent pipeline: invention described in INVENTION_BRIEF.md — CN, utility model
• 写专利：基于图神经网络的故障诊断方法，目标 EP

Work through the stages and confirm each one

The skill writes its outputs into a patent/ folder, including patent/INVENTION_BRIEF.md when your input is conversational. Progress is saved to patent/PATENT_STATE.json after each phase. If a session is interrupted, invoking the skill again within 24 hours resumes from the saved phase; older state is treated as stale and the run starts fresh.

By default the skill does not move on without your explicit confirmation. Review each output, especially the claims, before continuing.

• Drafts are written as markdown for review; docx is the option for filing-ready output.
• Set autonomous mode only if you explicitly request it.

Know the safety limits

The skill can run scripts, use web search and fetch, and send material to a Codex reviewer. If your invention is confidential, do not send sensitive text to external reviewer services or web search. Public disclosure before filing can harm patentability, so check this before running the pipeline on unpublished work.

The output is a draft. Legal accuracy, enablement and claim scope need review by a qualified patent attorney, and mistakes can permanently cost you rights.