How to Draft Patent Claims in 2027

IP Author
November 20, 2026
8 minute read
Generative AI improving patent quality and clarity

Quick summary: how to draft patent claims in five steps

  1. Put every element in the body. Keep the preamble short so it does not become a limitation.
  2. Link every function to structure. Replace words like module, unit and handler with processors, memory and a step-by-step algorithm.
  3. Anchor terms of degree. Give “about” and “substantially” a number, standard or example in the specification.
  4. Claim only what you can enable. Tier the claims from a mechanistic genus down to examples.
  5. Recite the technical mechanism. For software and AI claims, claim the component that improves the computer, not the business result.

Finish with the pre-filing check at the end of this guide, and rerun it after every amendment.

The majority of patent claims that fail to pass examination do not fail due to prior art but due to the language used in the claims. Using overly broad terms in the preamble, vague terms like “means” and undefined terms like “about” and “genus” in the claims will cause problems. Below are seven examples from the Federal Circuit from 2025-2026 showing what went wrong and how to fix it so you can see how to draft patent claims that hold up. 

This is a common issue. A study carried out by the USPTO Office of the Chief Economist found that over 1/3 of over 4 million office actions sent from 2008 to 2017 included a Section 112 rejection. Many of these will be down to language issues in the claims.

Did you know? In the USPTO’s research dataset of about 4.4 million Office actions mailed from 2008 to mid-2017, roughly 70 percent of Section 112 rejections were made under 112(b), the provision that requires claims to particularly point out and distinctly claim the invention. Source: USPTO Office of the Chief Economist, Office Action Research Dataset paper

Where claims are failing now

Case What failed The fix
Moskowitz v. Globus (2026) Body relied on preamble, so “universal” limited the claim Introduce elements in the body
Canatex v. Wellmatics (2025) “Second part” should have read “first part” Check antecedents per chain
Fintiv v. PayPal (2025) “Payment handler” read as means, no algorithm Tie functions to structure
Optis v. Apple (2025) “Selecting unit” was not structure Avoid nonce words
Enviro Tech v. Safe Foods (2026) “About” undefined after a range amendment Anchor terms of degree
Satius v. Samsung (2026) Claim scope not enabled Claim what you can enable
Recentive v. Fox (2025) Generic ML applied to new data Recite the technical mechanism

Canatex is the one outlier: the court corrected the error. But the patentee lost in the district court and had to win on appeal first.

Patent claim basics before the rules

Every claim is a single sentence with three parts. The preamble names the invention. The transitional phrase links the preamble to the body. The body lists the elements and how they relate. In the Rule 1 example below, “A system” is the preamble, “comprising” is the transitional phrase, and everything after the colon is the body.

The transitional phrase sets how open the claim is (MPEP 2111.03). “Comprising” is open-ended and does not exclude additional, unrecited elements. “Consisting of” excludes anything not specified in the claim. “Consisting essentially of” sits in the middle.

Claims also come in two types. An independent claim stands on its own and sets the broadest scope. A dependent claim refers back to an earlier claim and adds a limitation, which gives you a fallback position if the independent claim has to be narrowed. The rules below apply to both.

Rule 1: Include all elements in the body

In Moskowitz, terms in the body such as ‘intervertebral cage’ had an antecedent basis in the lengthy preamble. This was an indication to the court that the preamble was limiting. Therefore, universal became a limitation and the patentee admitted that they had not infringed. To avoid a limiting preamble keep the preamble short and ensure all structural elements are included in the body.

This also helps prevent any antecedent issues. Any elements that are introduced should be introduced with “a/an” and referenced with “the intervertebral cage.” Antecedent basis is where independent vs dependent claims most often go wrong, so ensure that all references are to one element within their own chain of dependency (MPEP 2173.05(e)). Examiners will have a stricter approach. If the meaning cannot be understood they will consider the claim to be unpatentable (In re Packard). Even if a court may be lenient.

The errors experienced drafters still make:

  • “A processor” in one limitation, “the controller” in the next
  • “At least one sensor” followed by “the sensor”
  • “A first valve” and “a second valve,” then “the valve”
  • A dependent claim that refers to an element introduced only in a sibling claim

Before:

A system comprising a plurality of sensors; a processor configured to receive data from the sensors and generate an alert if the measurement is greater than the threshold.

After:

A system comprising: a plurality of sensors, each configured to provide a measurement, and a processor configured to compare each measurement to a threshold and generate an alert if at least one measurement is greater than the threshold.

Ensure to check all chains after any amendments. The majority of antecedent issues in issued patents are down to limitations being moved to claim 1 with no antecedent.

Rule 2: Link all functions to structure

Since Williamson v. Citrix (2015), omitting means it no longer protects you. Fintiv held that payment handler triggered 112(f). Handler was compared to module, and configured to gave no structure. Optis also held that selecting unit triggered 112(f). Once a computer function is subject to 112(f) the algorithm must be disclosed in the spec; otherwise, your claim is indefinite (Aristocrat v IGT (2008)).

Before:

classification module configured to classify the image into one of a plurality of categories.

After:

one or more processors; and memory storing instructions that, when executed, cause the one or more processors to extract a feature vector from the image using a convolutional neural network, determine a similarity score between the feature vector and a centroid vector of each category, and assign the image to the category with the highest score.

One of the patent claim drafting best practices is to search all drafts for module/unit/handler/mechanism/device and replace with processor/memory or describe the functionality. Ensure to give a step-by-step algorithm within the specification. 

Rule 3: Link terms of degrees

When learning how to write patent claims, note that terms such as substantially and about are not indefinite in themselves but need to be objectively linked to something in the intrinsic record (MPEP 2173.05(b)). In Enviro Tech the term “about” was considered indefinite, as it was within a pH range. The claim had been amended to avoid prior art; however, there was no evidence in the record of how much variance was allowed. When moving a range to overcome a reference, be sure to include a definition of “about” in the record.

Before:

The coating has a substantially uniform thickness.

After:

The coating has a thickness that varies by no more than 10 percent of the mean thickness of the coating, as described in paragraph [0042].

Rule 4: Enable what you can claim

In Amgen v. Sanofi (2023), the Supreme Court held that claims to all antibodies that bind to a certain site and inhibit PCSK9 were not enabled. Even though 26 examples were disclosed. The further you go with your claims, the more you must enable. As seen in Satius, this applies to more than just biotech. The court held that the claims were not indefinite for including an impossible option. However, they were still invalidated due to lack of enablement.

Any claims that define the function of something will run into this issue (a model trained to…) when you draft patent claims. Ensure to tier your claims. Give a structural/mechanistic genus in claim 1, then smaller groups with common features in dependents and finally examples. Within the specification, describe what features are giving the function. 

Rule 5: List technical process

Section 101 is a two-track system. In the USPTO, Ex parte Desjardins (now precedential from Nov 2025) held that claims to the training of a machine learning model were eligible, as it improved the model. The USPTO have updated the MPEP (2106) accordingly. In court, Recentive v. Fox (2025) held that claims applying generic machine learning to new data, without improving the model, were ineligible. Desjardins is binding on the examiners. Therefore, draft to overcome the court.

Within the spec, ensure to define the technical problem and solution. Within the claims are all components contributing to the improvement. Ensure to highlight the benefit to the computer, not the business, for example memory usage or latency of the named process.

Before:

The scheduler is configured to optimize the allocation of compute resources.

After:

The scheduler is configured to allocate each job of a plurality of jobs to a node of a plurality of nodes, wherein the allocation is based on a predicted runtime of the job and available memory of the node, and reallocate a job from a first node to a second node if the available memory of the first node falls below a reserve threshold.

A pre-filing check

Rerun the first two after every amendment.

  • Every “the” or “said” has exactly one antecedent in its own dependency chain.
  • Every claim term appears in the specification with the same meaning.
  • Nobody term leans on the preamble, unless you want the preamble to limit.
  • “Module,” “unit,” and “handler” are tied to structure or backed by an algorithm.
  • Each term of degree has a number, standard or example in the specification.
  • The specification enables the full scope of each functional limitation.
  • Software and AI claims recite the mechanism the specification names as the improvement.

Things this guide cannot tell you

Some of the cases are from 2025 and 2026, so ensure to check if anything has happened since. The 112 rejection stats are from the USPTO using data from 2008-2017. This should not be used as a substitute for your own judgement on a particular invention.

Next steps

Many of the issues above are quite mechanical to spot. IP Author will help you by analyzing the antecedent basis, terminology, chains of dependency and enablement during patent claim drafting. This means you can spend your time focusing on the scope and strategy. Now that you know how to draft patent claims that hold up in examination and in court, give The IP Author a try with your next set of claims.

Keep reading

IP Author

5 Office Action Response Mistakes That Can Cost You Time, Deadlines, or Patent Rights

Nov 13, 2026
5 min read
Read article
Blog, Featured

IP Author for Claude: Patent Search, FTO and USPTO Examiner Data Inside Claude

Oct 30, 2026
6 min read
Read article
Blog

IP Author is now available in ChatGPT

Oct 19, 2026
7 min read
Read article
Start your 14-day
free trial.

Join 1,000+ patent professionals using IP Author

14-day free trial
Join 1,000+ patent professionals
Core workflows included
No obligation to continue
SOC 2 Type II · ISO 27001:2022