Tutorial · Patent workflow

Patent drafting API: search and claim workflow

A patent draft is an argument about scope, support and prior art. Use the model to retrieve, structure and check; leave the legal conclusions to a registered practitioner.

Short answer. Classify the disclosure into CPC candidates, run a documented prior-art search, build a claim tree with explicit support in the specification, and generate section drafts that cite the disclosure. Flag novelty and obviousness as risks for attorney review, never as conclusions.

Set the automation boundary

The model is good at recall, structure and consistency: finding similar language, expanding synonyms, checking antecedent basis, mapping limitations to paragraphs, and noticing that a dependent claim refers to a limitation no parent introduced. It is not qualified to decide novelty, written-description compliance or claim scope.

TaskAutomateHuman gate
ClassificationCandidate CPC classes and keywordsSearcher confirms the plan
Prior artQueries, filters, deduplication, family groupingReviewer selects references
ClaimsClaim tree, antecedent checks, support mapPractitioner sets scope
SpecificationSection drafts and figure descriptionsInventor confirms disclosure
NoveltyReference-to-limitation mapping and risk flagsLegal conclusion by counsel

Applications can contain unpublished invention data and export-controlled technology. Decide what may leave your environment before the first prompt; see legal AI data residency.

Start with the disclosure and extract the technical problem, components, method steps, materials, parameters and intended effect. Build a plan with synonyms, acronyms, broad and narrow terms, and classification filters. Search US and foreign patents plus standards, papers and product manuals.

Use the USPTO Open Data Portal for US patent data. The legacy PatentsView search host was retired during the 2026 migration, so use the current ODP endpoint and key. Record the endpoint, query, date, result count and references reviewed; the search log is part of the work product.

Deduplicate by patent family, not title. The same invention can appear in several family members. Rank by claim relevance.

Use CPC classes as a search aid

The Cooperative Patent Classification is a joint USPTO and EPO scheme. Its symbols describe technical subject matter, from broad to narrow. A class is a search aid, not a novelty opinion.

  • Classify the disclosure: ask for candidate classes at subclass and main-group level, with a rationale tied to a technical feature.
  • Walk the hierarchy: search the class, its parent and nearby siblings. A narrow subgroup can miss a document classified higher up.
  • Combine with keywords: classification catches vocabulary mismatch; keywords catch classification error.
  • Check the date: a class can be redefined. Use the version in force for the document's publication date when available.

Build the claim tree

A claim set is a tree. Each independent claim defines scope; each dependent claim adds a limitation and inherits its parent's scope. The tree shows where the fallbacks sit and which dependent claims lack support.

ClaimTypeDepends onAdded limitationSupport
1Independent—Sensor, controller and threshold¶[0012]–[0018], FIG. 2
2Dependent1Sensor is an inertial measurement unit¶[0021]
3Dependent2Threshold varies with temperature¶[0024]–[0026]
4Independent—Method of calibrating the sensor¶[0030]–[0036], FIG. 4

Ask for three mechanical checks: every dependent claim refers to a preceding claim, every limitation has antecedent basis, and every limitation maps to a disclosure paragraph or figure. These catch drafting errors, not validity or commercial scope.

import json
from openai import OpenAI

client = OpenAI(api_key="YOUR_KEY", base_url="https://api.plugsky.com/v1")
disclosure = open("disclosure.md", encoding="utf-8").read()
schema = {"cpc_candidates": [{"symbol": "string", "rationale": "string"}],
          "claim_tree": [{"number": 1, "type": "independent", "depends_on": None,
                          "limitations": ["string"],
                          "support": [{"paragraph": "string", "quote": "string"}]}],
          "support_gaps": ["string"]}
response = client.chat.completions.create(
    model="plugsky-pro", response_format={"type": "json_object"},
    messages=[{"role": "system", "content": "Draft support for a patent attorney. Return CPC "
               "candidates, a claim tree and a support map. Quote the disclosure for every support "
               "item. List unsupported limitations as gaps. Do not state that an invention is novel. Return JSON."},
              {"role": "user", "content": json.dumps({"schema": schema, "disclosure": disclosure})}])
draft = json.loads(response.choices[0].message.content)
for claim in draft["claim_tree"]:
    if claim["depends_on"]:
        assert claim["depends_on"] < claim["number"]
    assert claim["support"] or claim["number"] in draft["support_gaps"]

An unsupported claim must appear as a gap. In production, also check that each quoted string appears in the disclosure.

Draft the specification sections

A US non-provisional specification typically includes a title, cross-reference, background, summary, brief description of drawings, detailed description, claims and abstract. The legal anchor is 35 U.S.C. 112(a): written description, enablement and best mode. MPEP 2161 treats them separately.

Generate each section from the inventor's disclosure and cite the source paragraph. The background should not concede that a particular solution was known or obvious. The summary should not add a limitation absent from the detailed description. The detailed description should let a skilled person make and use the invention. The abstract should be technical, not a sales pitch.

Do not add technical facts the inventor did not provide. A fluent paragraph about a material, range or result nobody supplied is new matter. Mark missing details as open questions.

Novelty risk and the attorney gate

Novelty and obviousness are legal conclusions under 35 U.S.C. 102 and 103. The model can prepare a claim chart mapping each limitation to each reference, a list of differences, and references that disclose a feature in combination. A registered practitioner decides what those differences mean.

  • Claim chart: one row per limitation, one column per reference, with a quote and citation. Mark “not disclosed” explicitly.
  • Risk memo: closest references, differences, common features and questions for the inventor.
  • Attorney gate: no filing, inventor legal opinion or freedom-to-operate conclusion leaves without practitioner review.
  • Audit: keep the disclosure version, search log, model and prompt version, claim tree and reviewer notes.

This is not legal advice. Patentability, scope, infringement and validity require a registered patent practitioner and the specific facts.

Evaluate the workflow

Measure the parts separately. Track search recall on matters with known references, claim support coverage, drafting defects such as antecedent and dependency errors, attorney edit distance by section, and hours from disclosure to review and review to filing-ready.

Run the workflow on closed matters first. Compare the generated claim tree and search log with what the attorney filed.

FAQ

Can AI draft a patent application?

It can draft sections, propose a claim tree, classify the disclosure and assemble a prior-art search log. A registered patent practitioner must set claim scope, review the specification and make the legal conclusions. The model's output is a draft, not a filing.

How do I search prior art with an API?

Use the USPTO Open Data Portal for US patent data, combine CPC classification with keyword and synonym queries, include foreign patent documents and non-patent literature, and log every query and result count. Deduplicate by patent family.

What is a claim tree?

A claim tree shows each independent claim and the dependent claims that inherit from it, with the added limitation and the disclosure support for each limitation. It makes scope and fallback strategy easier to review.

Does the model decide novelty or obviousness?

No. It can map references to claim limitations and flag differences, but novelty and obviousness are legal conclusions under 35 U.S.C. 102 and 103. A registered practitioner must make them.

How do I keep invention data confidential?

Decide what may leave your environment, use a deployment with the retention and residency terms you need, restrict access by matter, and keep an audit log. Verify the current terms in the contract before sending unpublished invention data.

Sources: USPTO MPEP § 2161 on the three requirements of 35 U.S.C. 112(a), the Cooperative Patent Classification about page, and the USPTO Open Data Portal. All checked 20 November 2026. Reviewed 20 November 2026. This page is not legal advice; patentability and claim scope require a registered patent practitioner.