Ex parte Ziegler: A Split Panel on Which Technology You Improved

The application is about assembling context for code completion. The ordinary approach feeds a model the lines above your cursor. This one also gathers text after the cursor, text from open editor tabs, text from related files and file metadata, ranks it, builds a prompt from a tree structure of the surrounding code, and sends that to a code language model.

An examiner called it a mental process with nothing more. The Board disagreed, two to one, and the disagreement is the interesting part.

The Majority Decided a Narrower Question Than You Might Expect

Most of the majority opinion is about what the Examiner did not do, rather than about what the claim is.

Within Prong Two, which is where this was decided, the MPEP sets an order for the improvement inquiry. First look at the specification and ask whether a skilled reader would recognise an improvement. Second, check that the claim itself reflects it. The specification identifies the improvement; the claim has to embody the mechanism. The majority found the first step missing: "The Examiner has not provided an evaluation of the specification", and called what was offered instead conclusory.

So the majority did the step itself. It found the specification treats code editors and code completion as a technological field, and that the claim carries the improvement through specific steps: a tree structure built from the code before and after the cursor, sibling relationships, objects moved within that tree. Its holding sentence is that claim 1 "reflects a technological improvement in the functioning of code editors, not a generic invocation of hardware".

The Dissent Asks a Different Question

Judge Fenick would have affirmed, and the reasoning is not a quibble about the facts.

The Federal Circuit held in Recentive Analytics v. Fox that "patents that do no more than claim the application of generic machine learning to new data environments, without disclosing improvements to the machine learning models to be applied, are patent ineligible". The dissent reads this claim as exactly that, and treats Microsoft's own briefing as conceding it, since Microsoft argued the improvement was the use of unconventional input sources.

The dissent also cites a non-precedential 2026 decision for the same proposition, that "Generic use of AI without other parameters, such as improving the mathematical algorithm or making machine learning better, is abstract."

Put the two side by side and the split is clean. The majority asks whether a technology was improved and answers that the code editor is the technology. The dissent asks whether the model was improved and answers no. The majority's answer decided this appeal. Which question is the right one to ask in general is left exactly where it was.

The Same Claim, Read Twice

The last two rows are the ones to sit with. The majority never mentions Recentive. The dissent's central authority goes unaddressed in the opinion it dissents from, so the two are not arguing with each other so much as past each other.

Where It Sits Between Two Cases You May Already Know

• Ex parte Desjardins found eligibility where a claim limitation reflected an improvement to how the model itself operates.
• Recentive found ineligibility where generic machine learning met a new data environment with no model improvement.
• This claim changes neither the model nor its mathematics. What it adds is a claimed process for building the context: a tree structure from the code on both sides of the cursor, sibling relationships, objects moved within it. The majority read that process as improving the editor. The dissent read it as a new way of selecting input for an unchanged model. That is the gap between the two, and the panel split inside it.

That is a description of three decisions, not a prediction about a fourth. A split agency decision is not precedent and binds no examiner in another application.

This Is Not a Rule That Better Prompts Are Patentable

The easiest way to misread this case is as a holding that giving an AI model more or better context can be patented. The majority did not say that, and the claim it examined was not that.

What the majority credited was a combination: a stated shortcoming in existing code editors, a claimed process for constructing and organising code context, a specification explaining how that process changes how the editor behaves, and claim language that actually requires the process rather than describing the benefit. Take away the tree structure, the sibling relationships and the rearrangement, and what is left is the thing the dissent said it was all along.

The useful question this case raises is narrower and harder than "is better context patentable". It is when assembling context becomes an improvement to a technical system, rather than a new rule for choosing what to feed a generic model. The majority answered that for this appeal. It did not answer it generally, and a non-precedential split decision cannot.

What This Case Is Not

One drafting observation survives the split, because both opinions accept the order of the analysis even while disagreeing about its result. The specification is read first, and the claim is checked against it. A specification that never names its technical field, never says what the prior art could not do, and never explains why the change is an improvement leaves the second step with nothing to find.

Educational, not legal advice. Whether any particular claim is eligible under § 101 is a legal conclusion a qualified patent attorney reaches on the actual claims.